THE BLUEPRINT: 🗞️ News

AI For The Rest Of Us, Take Two

Sean 🤓 — Tue, 09 Jun 2026 08:02:05 GMT

Yesterday, Apple spent roughly half of the WWDC keynote talking about Apple Intelligence.

HALF.

That’s a lot of time for any one topic at an Apple keynote.

Two years ago, when Apple first introduced Apple Intelligence, I wrote that it felt like AI for the rest of us. Not because Apple had the smartest model, or because Siri was suddenly going to beat ChatGPT overnight, but because Apple seemed to be starting from a different place.

Apple promised a Siri that understood your personal context, protected your privacy, and worked inside the devices and apps you already use every day. Things didn’t pan out that way, and it’s obvious now that the technology of 2024 and even 2025 wasn’t capable of delivering against Apple’s lofty ambitions.

But in 2026 we’re in a completely different place, and Apple wouldn’t be putting Apple Intelligence front and centre again if it wasn’t 100% sure Siri AI was going to ship later this year.

Apple had the right idea all along

Looking back, I don’t think Apple’s original idea was wrong. If anything, I think it was more right than most of the AI products we’ve seen over the last two years.

The best version of AI was never going to be a blank text box that sits in a browser tab. It was always going to be something that understood what you were trying to do, had access to the right context, and could help you get things done inside the products you already use.

That was the promise of Apple Intelligence from the beginning. Personal context. Privacy. App Intents. On-device processing. A Siri that could understand not just what you said, but what you meant, what you were looking at, and what you were trying to do.

What’s different now is that the architecture looks much more capable of delivering against that vision. Apple has rebuilt the models, partnered with Google to use the technology behind Gemini, expanded Private Cloud Compute, and connected Siri more deeply into Spotlight, apps, screen awareness and personal context.

So the idea hasn’t really changed. It’s just finally starting to look technically possible with the technology of 2026.

This is the Siri Apple had to ship

For years, Siri has been useful for timers, weather, music, reminders and the occasional smart home command. But it never really felt like an assistant. It didn’t understand enough about you, couldn’t see enough of what you were doing, and couldn’t reliably take action across the apps where you actually get things done.

Siri AI finally looks like the product Apple was describing at WWDC two years ago. It can understand what’s on your screen, search across your messages, photos, emails and files, use up-to-date world knowledge, and take action through apps.

This is the Siri Apple had to ship because this is what an assistant really is. Not a voice interface. Not a chatbot. Not a smarter search box. Something that can understand what you’re trying to do, help you do it, and become the new interface for the Apple ecosystem.

Just ask for what you want

Apple is also making natural language a first-class citizen across its ecosystem.

Creating a calendar event by describing it is useful. Asking Siri to find the right photos and add them to a shared album is useful. Comparing files on your Mac and asking which one solves a specific problem is useful.

But the more powerful and interesting examples were the technical ones. Building a Shortcut by describing what you want. Creating a custom Safari extension in natural language. Asking Siri to keep an eye on a website for when it’s updated.

These are things most people either don’t know how to do, or don’t even realise are possible. That’s what makes this exciting. This is what AI should be doing: helping people get more out of the technology they already use, without needing them to understand how any of it works.

What this means for brands

There is also a bigger implication for brands and marketers.

For the last year, most of the discussion around GEO has focused on ChatGPT and Gemini. That makes sense, but I think Siri AI now needs to be part of that conversation too.

If Siri AI becomes a new interface for the Apple ecosystem, it also becomes a new way for people to discover information, compare options, make decisions and take action. That means brands will need to understand how they show up in Siri AI not just ChatGPT and Gemini.

But there are some big unknowns. Nobody really knows yet where Siri AI will get all of its knowledge from, how much will come from Gemini, how much will come from Apple’s own systems, or whether showing up well in Gemini means you show up well in Siri AI.

The more significant difference is apps. Until now, a brand’s app has had very little influence on how it shows up in AI answers. Siri AI could change that because Apple is giving developers specific ways to make app content and actions available to Siri, Apple Intelligence, Spotlight, Visual Intelligence and Shortcuts.

That means App Intents, in-app search, Spotlight indexing, structured app content, product feeds, location data and APIs all start to matter much, much more. They will help Siri AI understand what your brand does, what you sell, where you are, what actions people can take, and when you are relevant.

The Apple difference

This is why Apple’s approach to AI still feels different.

Most AI companies are trying to pull us into their products. Open ChatGPT. Open Gemini. Open Claude. Ask a question. Write a prompt. Start a new conversation.

Apple is trying to do something else. It is trying to put AI inside the products people already use, so your phone, Mac, Watch, apps, camera, messages, files and photos all become more useful.

That’s a much more Apple-shaped version of AI. Less about the model. Less about the prompt. More about the experience.

There is one annoying caveat, which is that Siri AI won’t initially be available in the EU on iOS and iPadOS. So I may not even be able to properly test the thing I’m writing about straight away.

But that doesn’t change the bigger point. Apple is showing us what personal AI could be. Not a chatbot bolted onto your life, but intelligence built into the technology you already use.

Sources: Apple | WWDC Live Stream | The Verge | TechCrunch

A week in Generative AI: Knowledge Work, Self Improvement & Sovereign Wealth

Sean 🤓 — Sun, 07 Jun 2026 18:01:19 GMT

This week in AI news, OpenAI pushed Codex further beyond software development and into broader knowledge work, claiming that around 20% of users are already using it this way. Anthropic also published an update on its progress towards AI systems that can help build better versions of themselves, alongside a warning that frontier AI development may need coordinated slowdowns if things keep accelerating. And Bernie Sanders proposed the American AI Sovereign Wealth Fund Act, which would give US citizens a stake in major AI companies. It is an interesting idea, but also an awkwardly national answer to a global problem, given that the data AI models have been trained on comes from the world, not just America.

In Web 4.0 news there’s a great report from Harvard Business Review on How People Are Really Using AI in 2026 and in Ethics News UK media groups were given the power to opt out of Google AI search summaries, which is a big deal.

I’d also be really grateful if you could take 2 minutes to complete a quick reader poll about The Blueprint.

Enjoy!

OpenAI launches new Codex tools for white-collar work

OpenAI is starting to reposition Codex from a coding agent into a broader agent for knowledge work. The company says Codex now has more than 5 million weekly active users, and that knowledge workers now make up around 20% of its user base.

This feels significant because the coding-agent pattern is clearly starting to escape software engineering. First developers got agents that could plan, execute and ship work across files, tools and workflows; now the same pattern is being applied to reports, spreadsheets, presentations, contracts, research and data analysis.

Coding is becoming the test case for the future of work. If agents can handle software development, which is complex, messy and full of edge cases, it is not hard to see how the same model moves into strategy, finance, marketing and operations next.

TechCrunch | OpenAI

Anthropic (Sorta) Calls for Pause on AI Development. You Should (Sorta) Take It Seriously

Anthropic has published an interesting new paper on how AI systems could eventually become capable of building better versions of themselves. We’re not there yet, but it argues that AI is already accelerating the development of AI.

Some of the numbers are pretty striking. Anthropic says more than 80% of the code merged into its own codebase is now authored by Claude, and that its typical engineer was merging 8x as much code per day in Q2 2026 as they were in 2024.

This is why Anthropic is now making the case for coordinated slowdowns if frontier AI development starts moving too quickly. Not a unilateral pause, which would simply hand advantage to less cautious players, but a verified, collective slowdown between the major labs and governments.

It feels like the near-term and long-term AI debates are starting to blur. We keep talking about AI replacing knowledge work, but the more significant issue may be AI accelerating AI itself, because once the development curve stops being purely human-paced, governance becomes even harder than it already is.

Gizmodo | Anthropic

Introducing the American AI Sovereign Wealth Fund Act

This week, Bernie Sanders put forward one of the more interesting AI policy ideas I’ve seen for a while. If AI has been trained on the collective knowledge, creativity and labour of humanity, then humanity should share in the wealth it creates.

The American AI Sovereign Wealth Fund Act would give the American public a direct ownership stake in major AI companies through a one-time 50% tax paid in stock rather than profits. It is a bold idea, but also an awkwardly national answer to a fundamentally global problem.

Because if AI has been trained on the work, data and culture of people all over the world, why should only American citizens share in the upside? That doesn’t make the underlying principle wrong, but it does expose one of the biggest tensions in AI policy: the technology is global, the companies are mostly American, and the political solutions are still largely national.

YouTube

Web 4.0

AI Ethics News

Long Reads

One Useful Thing - Co-Existence and the End of Co-Intelligence
The Guardian - My year with the robots: how Joanna Stern let AI into her home, work – and heart

A week in Generative AI: Opus, Valuations & Football

Sean 🤓 — Sun, 31 May 2026 18:01:09 GMT

Anthropic releases Opus 4.8 with new ‘dynamic workflow’ tool

Anthropic released the latest version of their frontier model, Opus 4.8, just 41 days after the release of the previous version. That release showed a huge jump in performance and when I wrote about it I suspected that it was actually more like a preview of Opus 5.

Opus 4.8 is more of an incremental improvement on Opus 4.7 but the killer feature is how it manages mistakes - it’s far more likely to flag uncertainty about its work and less likely to make unsupported claims. Both of these things have been a big issue for frontier AI models, and are particularly problematic for models that power agentic workflows where you need self-reflection and honesty.

As part of the release, Anthropic also commented on Mythos, their incredibly powerful cyber-security model and it sounds like it’s nearing general release. They’ve been making good progress on the required safeguards and expect to be able to release it to all their customers in the coming weeks. If that’s the case I suspect we will be seeing Claude Opus/Sonnet 5 towards the end of the summer.

Anthropic | TechCrunch

Anthropic reaches valuation of $965bn, beating OpenAI to become world’s most valuable AI firm

The good thing about funding rounds is that it puts an agreed valuation on a private company. In Anthropic’s recent round of funding they raised $65bn at a valuation of $965bn. This is the 8th funding round Anthropic have completed, with the first one on May 2021 raising ‘only’ $124m at a valuation of $845m, meaning they have over 10x their valuation in just 5 years.

Anthropic have been on an unbelievable growth tear over the last 6 months. Whilst Anthropic’s revenue is rarely publicly declared The Information has reported that Anthropic are now generating an annual revenue run rate of $45bn, up from roughly $1bn 18 months ago. In Q1 2026 Anthropic predicted their revenue would grow 10x, but it in fact grew 80x - madness!

As Simon Willison has written, Anthropic and OpenAI have found product-market fit which is what’s driving these revenues. That fit is coding agents with enterprise customers paying huge amounts for API use. The medium term bet is that they will be able to extend this success beyond coding agents to general purpose knowledge work agents and that’s when usage (and revenue) will really explode.

If Anthropic and OpenAI are able to crack that then 80x revenue growth in a single quarter will look like pocket change. I think we’re in for a bit of a wild ride over the next couple of years!

The Guardian | The Information | Fortune | Simon Willison

Are consumers ready for humanoid robots?

Great consumer report from the FT on the current state of humanoid robotics and where it’s headed. Interesting stats they shared:

Investment has tripled over the last 2 years to over $40bn
Shipments are expected to grow from 90k to 1.2m in 2030
The market will be worth $200bn by 2035

There’s still a long way to go but a lot of investors and start ups are betting on being able to prove that an affordable general-purpose humanoid robot will be the next big consumer tech product.

YouTube

School of Football

Great promo for Hyundai (who own Boston Dynamics)’s sponsorship of this year’s World Cup. I not for a second suggesting that a robot can learn how to play football just by watching some footage, but it will be fun to see how they get on teaching Atlas to play!

YouTube

Web 4.0

AI Ethics News

Long Reads

One Useful Thing - Choosing to Stay Human
Benedict Evans - Predicting AI job exposure
Stratechery - The SpaceX IPO and Data Centers in Space

A week in Generative AI: Google, Google & Google

Sean 🤓 — Sun, 24 May 2026 18:01:03 GMT

This week all the major AI news came out of Google’s I/O 2026, where they announced a huge range of new AI models and AI-powered products and capabilities. There was a lot to take in so I’ve covered everything I thought was important in this week’s newsletter. There was also a great new video from Boston Dynamics showing off Atlas’ ability to lift a fridge!

As you can imagine, there was lots of Google news that related to Web 4.0 as well, with the most notable one being their new Universal Cart that will follow you across the internet.

There was lots of Ethics News too and it’s worth checking out the two Long Reads this week as well with a great article from Simon Willison on The last six months in LLMs.

Was there anything at Google I/O 2026 that didn’t include AI?!

Google held its annual I/O event this week and everything it announced was related to AI in one way or another. There was a huge amount of new things that they covered, which was difficult to follow at times, but The Verge has a great summary of the 13 biggest announcements at Google I/O 2026 you can check out if you’re interested in the headlines.

The big things I’d highlight are:

The announcement of A new era for AI Search that I cover below in more detail
The release of a new family of models, Gemini Omni that I go into in more detail below
The release of Gemini 3.5 Flash, which is focused on powering AI agents, that I go into more detail below
Gemini Spark, Google’s take on OpenClaw, which is coming soon
An upgrade to the Genie world model which can now simulate real streets with Street View

I’m sure we’re be learning a lot more about all of this over the coming weeks as things are released and people get to go hands on. A lot was covered at I/O so I’ll do my best to cover the important things in more detail below!

The Verge

A new era for AI Search

At I/O, Google shared a slick marketing video positioning its latest Search announcements as the beginning of a new era for AI search. They clearly want Search to become more conversational, more agentic, more personal, and less dependent on the traditional list of blue links.

But the reality isn’t quite as dramatic. The biggest live change is an upgrade to the model powering AI Mode, which most people probably won’t notice, while the more ambitious features, including search agents, agentic booking, generative UI, and mini-apps, are either coming this summer or rolling out first in the US and/or to paid Google AI subscribers. So this is unlikely to transform the everyday search experience for most people overnight.

The important distinction is that most of these changes are coming to AI Mode, not necessarily to the default Google Search experience. AI Mode is Google’s dedicated conversational search interface, and while Google is clearly creating more routes into it from standard Search, that is not the same as replacing the core search experience for everyone.

YouTube | Google

Gemini Omni will create anything from anything

Gemini Omni follows the same direction of travel we’re seeing across AI image, video and coding tools: combining reasoning with generation so the model can understand intent, not just produce an output. With Omni, you can combine images, audio, video and text as inputs to generate videos grounded in Gemini’s real-world knowledge.

For example, you can take an existing video and ask Omni to change a specific part of it, like turning a sculpture into bubbles, making a mirror ripple like liquid when someone touches it, or transporting a violinist into a completely different visual environment while keeping the character and action consistent. Users can then iterate, with each new instruction building on the last.

According to The Verge, Google’s new anything-to-anything AI model is wild

The first version of Gemini Omni, the Flash model, is now available to Google AI Plus, Pro and Ultra subscribers. It can currently create 10-second videos, with image and audio outputs still to come.

Content generation is starting to shift from prompting to directing, in the same way coding has shifted recently from writing code to managing coding agents. Instead of describing a finished asset from scratch, you can bring Gemini a set of references, existing assets and half-formed ideas, then shape the output through conversation. That feels much closer to how AI-assisted creative tools should work.

YouTube | TechCrunch | Google

With Gemini 3.5 Flash, Google bets its next AI wave on agents, not chatbots

Gemini 3.5 Flash is Google’s latest flagship model, and the company is now explicitly positioning it around agentic work: coding, tool use, long-running workflows, subagents and enterprise tasks that need speed as much as intelligence.

That focus on speed feels important. I use coding agents a lot, and latency is still one of the biggest areas where the experience needs to improve. So far, most of the conversation around frontier AI models has focused on intelligence and cost, but agents add a third requirement: speed.

That matters because agents do not just answer a question once. They reason, call tools, write code, inspect results, make changes, run checks and sometimes split work across multiple subagents. Once models are doing that kind of multi-step work, waiting for the next token becomes part of the product experience.

TechCrunch | Google

Atlas, can you bring me a drink?

Great new Atlas video from Boston Dynamics that shows Atlas picking up a whole fridge. This is the first time I’ve seen a humanoid robot use its whole body to lift something, like people do when lifting heavy objects. There’s a nice behind the scenes video of how they trained it to do this here if you’re interested.

YouTube

Web 4.0

AI Ethics News

Long Reads

Simon Willison - The last six months in LLMs in five minutes
Person Familiar - AI, “Humanity”, and Dr. Manhattan Syndrome

A week in Generative AI: Daybreak, Mouse Pointers & GD01

Sean 🤓 — Sun, 17 May 2026 18:01:02 GMT

No big AI news this week, although OpenAI announced Daybreak, their semi-answer to Anthropic’s Mythos model and Glasswing initiative. There was also a great video from Google DeepMind on putting AI into a mouse cursor (bear with me on this one!) and a ridiculous exoskeleton robotics video from Unitree.

In Web 4.0 news, OpenAI have launched ChatGPT for personal finance, just in case you needed that (?!) and in Ethics news it was reported that data centres now account for 6% of electricity supply in UK and US.

OpenAI just released its answer to Claude Mythos

So, OpenAI’s Daybreak isn’t a model, but more a way for cyber-security experts and large technology organisations to do similar security testing work as Anthropic are doing with their Mythos model.

This is quite different I think. Anthropic built a new model (Mythos) that they realised was incredibly capable to the point where it could be dangerous to put into consumer’s hands. Mythos was specifically very capable at coding and identifying/exploiting/fixing cybersecurity vulnerabilities sp Anthropic decided to make it available to a small number of organisations so they could get ahead of the cybersecurity threat that Mythos would pose when they release it.

OpenAI’s approach isn’t based on them building a new model that they think is incredibly capable/dangerous. They’re just trying to give organisations similar cybersecurity capabilities using their existing models.

This difference doesn’t matter now, but will matter when (and I do think it’s when, not if) Anthropic decide to release Mythos to the general public. There’s a chance that OpenAI are then behind the frontier for the first time as Anthropic will have a model more capable that anything else available.

Both Google and Anthropic have been behind the frontier of AI capabilities at certain points over the last 3.5 years, but generally OpenAI haven’t been. They’ve always been either at the frontier or pushing the frontier forward.

It’ll be interesting to see how they react to being behind the frontier for the first time.

The Verge

Reimagining a 50-year-old interface (the mouse pointer) with AI

This is one of those things that when you think about it is incredibly simple, but I am sure was technically very challenging. The phrase I like best from the video is how you ‘share attention’ with an AI model - essentially, how do you make sure an AI model knows what you’re looking at or referring to when you ask it to do something.

If you’d asked me before seeing this video if it would be a good idea to put AI ‘into’ a mouse cursor I would probably have laughed, but now I‘ve seen this video and I’ve though about it, the more I think it’s a great idea and part of the future of AI-powered users interfaces.

Love it!

YouTube

Unitree unveils GD01

This is a very weird, ridiculous video but I had to share it because it was funny to watch, but also it’s inevitable that something like this is actually in our future.

So, Unitree have been pushing the limits of what’s possible with humanoid robotics and made a lot of very rapid progress. This however, is not that.

The video shows an ‘exoskeleton’ that in theory would have a human operator inside, but I can’t see any controls and the part of the video where it’s moving with someone in it is clearly a dummy and not a real person. I also don’t really see how a human operator could operate it when it’s on all fours as they wouldn’t be able to see where they’re going!

YouTube

Web 4.0

AI Ethics News

Long Reads

Daring Fireball - AI Is Technology, Not a Product
Stratechery - The Inference Shift

A week in Generative AI: Revenue, Mythos & Helix

Sean 🤓 — Sun, 10 May 2026 18:00:46 GMT

No major new releases this week, but OpenAI did release GPT-5.5 Instant, which is a small bump rom GPT-5.3 Instant and a good upgrade for free users. They also released new realtime voice models in the API.

In other news this week, Anthropic was reported to have grown its revenue 80x in the last quarter (🤯). Mozilla also released a paper showing the huge number of bugs they fixed in the last month with the help of Anthropic’s Mythos model showing that there is a lot of substance behind the hype. And two Figure Helix robots made a bed 🤷‍♂️.

I was also disappointed on Saturday night to see a tweet from Sam Altman, that as an autistic person, I found offensive. He and OpenAI should be better than this, especially given how many neurodivergent people work in AI and technology.

Anthropic weighs deal for near $1tn valuation as revenue surges

This week it was reported that Anthropic’s revenue grew 80x in a single quarter, which is ridiculous levels of growth! To deal with all the new demand they’ve had to rent a whole data centre from xAI, bringing into question whether Elon Musk is pivoting the company to be a data centre builder, rather than a frontier AI company.

A lot have people have doubted the financial viability of the levels of investment that we’ve seen over the last 18 months into AI labs, data centres and chip manufacturing but I think there’s still a lot more demand to come and it’s unlikely we’re in full bubble territory yet. I might be wrong, but 80x revenue growth in one quarter is unprecedented and tells an interesting story.

FT | Gizmodo

How Anthropic’s Mythos has rewritten Firefox’s approach to cybersecurity

There was some skepticism that when Anthropic announced Mythos (their ‘too-powerful-to-release’ model) that it was just a clever marketing ploy. I didn’t buy that, and it’s good to now start getting an idea of the impact that it’s having for the limited number of organisations that have access to it.

Security researchers for Mozilla’s Firefox browser fixed 423 bugs in April, compared to the usual 20-30 bugs they fix every month. That’s astounding! Some of these bugs had lain dormant in the code for more than a decade.

And this isn’t automating or removing the need for very experienced and capable software engineers. Mozilla say that this is create more, higher value work, so they can improve the stability and quality of their web browser. It’s a great example of how AI is creating new opportunities for work and not just automating jobs away.

TechCrunch | Mozilla

Helix 02 Bedroom Tidy

Good to see Figure getting their sales pitch in early that you need two of their robots and not just one 🤣.

In all seriousness, it’s a bit weird seeing two robots work together and pausing to look at each other before performing an action together. I guess it’s something we’re going to have to get used to seeing though!

YouTube

Web 4.0

AI Ethics News

I was disappointed to see this tweet from Sam Altman. As an autistic person, I found it pretty offensive, and I suspect many other people did too.

Describing someone as an “autistic genius with very strange taste” leans heavily on tired stereotypes about autistic people being eccentric savants or socially odd caricatures. Even if intended jokingly, it’s a strange thing for the leader of one of the world’s largest AI companies to post publicly, especially given how many neurodivergent people work across the technology industry.

A week in Generative AI: Images 2.0, AGI & Robotics

Sean 🤓 — Sun, 03 May 2026 18:01:19 GMT

Not a big news week this week, but some interesting reflections on ChatGPT Images 2.0, and some good videos from Sequoia Capital’s AI Ascent Conference that include an interview with Demis Hassabis and a presentation from Jim Fan.

In Web 4.0 news there are rumours that OpenAI’s growth in terms of users and revenue is stalling (which might explain their pivot away from consumer and towards enterprise) and in Ethics News there’s some interesting research on Dead Internet Theory and signs that despite faster/higher adoption amongst young people that there’s the start of a backlash.

Enjoy!

Images 2.0 is on the Path to AGI

So, I have to admit that I didn’t realise how big a deal ChatGPT Images 2.0 was last week when it was first announced and I still don’t think many people have. It’s the ‘new thinking’ capabilities’ that OpenAI have added to this model that people are starting to realise are really powerful.

Take the example above, Images 2.0 was able to create an image of a lego set, but also the instructions and parts list.

It’s good for eduction too, with up-to-date, factually correct information:

What makes it great isn’t the fact that it can generate images, or that it can generate lots of related images. It’s that it can go off, plan, do some research, fact-find and then string together the visual expression it needs to answer the prompt. All in one-shot.

Latent Space have a great article about how this put’s the ‘G’ in AGI that I think is worth a read to understand why image generating capabilities like these are important and not just a distraction.

Latent Space | Twitter

Demis Hassabis: We’re Three Quarters of the Way to AGI

Speaking of AGI, there was a great interview this week with Demis Hassabis as part of Sequoia Capital’s AI Ascent 2026 conference. He talked about where he started in game development and using early AI in his games, key startup lessons he’s learned, and how he believes AI is the ultimate scientific tool.

Demis believes that there is a lot more to come from the scientific application of AI, especially in biology and drug discovery. This builds on the work they’ve been doing with AlphaFold and he expects big advancements in the next couple of years.

The interview starts to veer into philosophy towards the end (which I like, but I know isn’t everyone bag). But it’s well worth some of your time to watch if you’re curious about the impact AI could have on humanity over the next 5 years.

YouTube

Robotics’ End Game: Nvidia’s Jim Fan

Another great presentation from AI Ascent was from Jim Fan from NVIDIA on why robotics may be about to have its own “GPT moment”.

His core argument is that robotics is starting to copy the playbook that made language models work: large-scale pre-training, fine-tuning, reinforcement learning and scaling laws. But instead of predicting the next word, robots need to predict the next physical state of the world.

He argues that video models, egocentric human data, simulation and “world action models” could finally solve the data bottleneck that has held robotics back. It gets pretty technical in places, but the big idea is simple: we may be moving from robots that are carefully programmed to robots that learn from watching, simulating and acting in the world.

Definitely worth watching if you’re interested in what happens when AI moves from pixels and words into atoms.

YouTube

Web 4.0

AI Ethics News

Long Reads

OpenAI - Where the goblins came from

A week in Generative AI: GPT-5.5, DeepSeek & Images 2.0

Sean 🤓 — Sun, 26 Apr 2026 18:01:31 GMT

Big week this week with the release of GPT-5.5, DeepSeek v4, and ChatGPT Images 2.0. There’s also a great video of a robot playing table tennis because why not?!

In Web 4.0 news, OpenAI has started testing cost-per-click ads and Anthropic is connecting Claude to lots more tools.

In Ethics news, there’s more commentary on Anthropic’s Mythos model, and the Pentagon is looking to pivot to AI-powered warfare.

A few great Long Reads to check out too - Nilay Patel writes about ‘Software Brain’ on The Verge which explores why some people are bullish on AI whilst others hate it. There’s also a great video from Gary Stevenson on whether AI will drive wages up or down.

OpenAI says its new GPT-5.5 model is more efficient and better at coding

The big release of the week was from OpenAI who made GPT-5.5 available to the world, hot off the back of Anthropic releasing Opus 4.7 last week.

The promotional video definitely positions GPT-5.5 around tool use and automation and shows off how the Codex app is a much more capable interface than the usual way of interacting with ChatGPT. I suspect we’re seeing a move towards the ChatGPT interface just being for ‘chatting’ whereas the Codex app is starting to pivot away from just being a ‘coding’ app to more of a general purpose ‘work’ app for getting things done.

As Ethan Mollick says in his review, it’s difficult to articulate how GPT-5.5 improves on previous models as AI still has a jagged frontier. But it does intuitively feel better, faster, and more efficient. It’s definitely progress, but it’s hard to articulate on what exactly.

OpenAI | The Verge | Ethan Mollick | Simon Willison

DeepSeek previews new AI model that ‘closes the gap’ with frontier models

It’s been just over 15 months since DeepSeek released their v3.2 and R1 models which temporarily sent tech stocks in the US spiralling. This week the released both DeepSeek V4 Flash and V4 Pro which are mixture-of-experts models with context windows of 1 million tokens each.

DeepSeek says they have almost “closed the gap” with current leading models, but both their models are text only. However, the models are much more affordable than any other frontier model available today and are available as open-weights models unlike any of their peers.

HuggingFace | TechCrunch | Simon Willison

OpenAI’s updated image generator can now pull information from the web

On Tuesday, OpenAI started rolling out its second generation image generating model, citing ‘new thinking capabilities’, higher resolution images and more consistency. It wouldn’t surprise me if we’re about to see (hear!) a new voice generating model too, as I’m sure that’s what they used for the narration of this video.

ChatGPT Images 2.0 certainly looks very impressive, but we haven’t seen the same Ghibli-style hype we saw with the first version a year ago. To put it to the test I gave it the last 6 weeks of newsletters I’ve written and a couple of the more recent longer reads I’ve published and it came up with the hero image for this week’s newsletter. You can judge for yourselves!

Simon Willison had a fun test of the new model too, getting it to generate Where’s Wally/Waldo style images, which I think he quickly regretted!

OpenAI | The Verge | Simon Willison

AI-powered robot beats elite table tennis players

Sony AI in Zurich have developed a table tennis playing robot, called Ace, that was able to win three out of five matches against elite players, but lost the two it played against professional players. This might be the Garry Kasparov moment for robotics.

Back in 1996, IBM’s Deep Blue, a specialised chess playing supercomputer, became the first machine to win a chess game against a reigning world champion. Similarly, Ace is a specialised table tennis playing robot - it uses multiple camera, specific ball-tracking and, isn’t bi-pedal, using an eight-jointed arm on a movable base.

It even pulled off one rapid backspin shot that a professional had thought impossible:

So, we’re still a waySteam way off a general humanoid robot being able to reliably beat table tennis professionals. It took about 20 years to get from Deep Blue to Alpha Zero, which was more of a general purpose game-learning system. I wonder how long it will take us to get to a general purpose game-playing humanoid robot?

The Guardian | Nature

Web 4.0

AI Ethics News

Long Reads

The Verge - Beware Software Brain
Garys Economics - Will AI drive wages down… or up?
The Ezra Klein Show - Stewart Brand on LSD, AI Black Boxes and the Beauty of Care

A week in Generative AI: Opus, Flint & Design

Sean 🤓 — Sun, 19 Apr 2026 18:01:46 GMT

Big week this week with the release of Claude Opus 4.7, which has been slept on by the mainstream media, the release of a genuinely creative AI model from a small Australian startup, Springboards and signs that Design might be the next area to be widely disrupted by AI. There was also a great TED talk from Peter Steinberger who shares his OpenClaw story.

Enjoy!

Anthropic Releases Claude Opus 4.7 to Remind Everyone How Great Mythos Is

On Thursday Anthropic released the latest version of their most capable Opus model, bumping it up to version 4.7. For some reason this wasn’t that widely reported on, but I think this is a bigger deal than most people think.

Firstly, Anthropic don’t release their Opus models very often. For example, the Opus models jumped directly from v3 to v4 whilst Sonnet (their mid-tier model) had v3.5 and v3.7 between major model releases.

Secondly, Opus has never before been on a higher version number than Sonnet. Ever. Anthropic have always been very deliberate and careful with their release strategy to roll out Sonnet first, and then if everything goes well from a safety perspective to then release the Opus model.

Thirdly, and probably most importantly, Opus 4.7 is a huge leap in performance that we don’t usually see with point releases, especially in just 2 months (Opus 4.6 was released in February). As you can see below, on agentic coding tasks, Opus 4.7 is a clear 10-15 percentage points more capable which is huge!

For me, all of this points towards Opus 4.7 being more like a preview of Opus 5, rather than just a point release improvement over Opus 4.6. This is a distilled version of Anthropic’s new Mythos model dressed up as Opus. Anthropic hint at this in their release, but don’t go as far as fully confirming it:

I’ve been putting Opus 4.7 through its paces with various coding and other general tasks over the weekend, and I have to say I’m very impressed. It’s not just coding tasks it’s materially better at, it’s also more thoughtful, considered, reliable, and thorough in all the tasks I’ve thrown at it.

I’m sure we’ll be seeing a lot more coverage and commentary of Opus 4.7 next week because it’s just too good to ignore.

Anthropic | Gizmodo

Flint by Springboards

I’ve been following Springboards for a couple of years now - they’re a great Australian start-up that are focused on marketing and creative work and building AI tools that enhance creativity instead of replacing it.

This week they released their first AI model, Flint, which has been trained specifically for inspiration and creativity. It’s designed to tackle a problem with our current AI models that not many people are aware of - they ‘collapse’ to the average, are brilliantly mediocre, and cater for the masses. This is all like poison for a creative process and Springboards want to address this head on with Flint.

The demo above is really worth a watch - it clearly shows how leading models from Anthropic and OpenAI keep giving repeating and predictable answers whereas Flint injects much more variation and creativity. Flint is in alpha preview now, and focused on classic marketing use cases, but I think has a lot of potential for broader creative tasks and workflows in the future. Flint will be great for anything that requires a spark of creativity. One to keep an eye on!

Springboards

Anthropic Launches Claude Design

Releasing their most capable model just wasn’t enough for Anthropic this week, so they decided to release a whole new capability in Claude Design. This is essentially a frontend UI design tool, but can be applied to lots of other things too like graphics and presentation slides.

This has the potential to be very disruptive and Mike Krieger, Anthropic’s CPO, even had to leave Figma’s board of directors as part of releasing this new capability. Following the release, Figma’s stock price also fell by 7%.

It’s an interesting time in the design space right now with Claude Design, and the launch of Google’s Stitch a few weeks ago. I think this is the next area to see major disruption after the huge amount of disruption we’ve been seeing in the coding space over the last 6 months. Because AI models are now so good at code, and frontend design (and other types of design) can be driven by code, they’re now very quickly becoming really good at design too.

As I’ve said before, software ate the world and now AI is eating software, taking the world with it.

Anthropic | TechCrunch | Gizmodo

How I created OpenClaw

OpenClaw has had a lot of hype this year, and deservedly so - it’s been the fastest growing open source project of all time and got a huge amount of love not just from the developer community but lots of press coverage too.

So it’s great to see the personal story behind the code in this TED talk from Peter Steinberger, who has recently joined OpenAI. He’s been on a wild journey over the last 6 months and regardless of what you think of OpenClaw it has been incredibly disruptive and has a huge impact on they types of tools AI companies are developing, and how they’re developing them.

TED

Web 4.0

AI Ethics News

Long Reads

Stratechery - Mythos, Muse, and the Opportunity Cost of Compute

A week in Generative AI: Glasswing, Muse & Affordability

Sean 🤓 — Sun, 12 Apr 2026 18:01:21 GMT

Three big news stories this week - Anthropic announced that they won’t release their newest model, Mythos, as it poses a major cybersecurity risk, Meta get back in the AI game, and Unitree announce pre-orders for their R1 robot outside of China for under $5k.

In Ethics News, a 20-year-old was arrested for allegedly throwing a Molotov cocktail at Sam Altman’s house, OpenAI shared their vision for the AI economy, and the US summoned bank bosses to discuss cyber risks posed by Anthropic’s latest AI model.

Good long reads from Helen Toner on AGI and the New Yorker on Sam Altman worth checking out too.

Anthropic debuts preview of powerful new AI model Mythos in new cybersecurity initiative

The frontier AI companies have been hinting at this for a while now - they’re going to get to the point where they train a model that has capabilities they’re hesitant to release to the general public. That moment has now come and the reason is cybersecurity.

This week Anthropic announced a new model, Mythos, that is so capable at coding that it poses serious cybersecurity risks. Instead of releasing it, they are partnering with some of the biggest technology companies in the world to test it and use it to identify security gaps in their own platforms before they decide whether to release Mythos more broadly. They’re effectively giving all the large technology platforms a ‘head start’.

To try and put this in context, Mythos found a bug that had gone undiscovered for 27 years. It’s also helped one cybersecurity professional discover more bugs in 2 weeks than they have throughout their entire career.

There is some scepticism that all of this is essentially ‘marketing cover’ to put more advanced models behind ever larger enterprise contracts but I don’t buy that argument. Although, I am a big cynical that complete co-incidence OpenAI also announced a few days later that they’re about release a new cybersecurity product.

What I think this shows is that we’re starting to reach a tipping point where models are so advanced that just releasing them isn’t a good idea. We’re now entering a period where frontier AI companies are going to have to think more carefully about how and when they release their most capable models.

Anthropic | TechCrunch | Axios

Meta debuts the Muse Spark model in a ‘ground-up overhaul’ of its AI

A month ago I wrote about Meta reportedly delaying the release of its latest model, but this week they did actually release it under the name Muse Spark. This is obviously Meta playing catch-up, so don’t expect it to introduce any new groundbreaking capabilities.

Simon Willison took it for a spin and noted that it has some interesting tools, and the release helped the Meta AI app climb to No. 5 on the App Store.

I’d file this as a stepping-stone release while Meta plays catch-up and they’ve promised that bigger more advanced models are to come and that they’re going to get back into the open-weights model game, which can only be a good thing.

Meta | TechCrunch | Simon Willison

Here’s The Most Affordable Humanoid Robot You Can Buy Now

Back in July last year, I shared that Chinese robotics company Unitree was releasing their R1 humanoid robot for less than $6k. They’ve just opened pre-orders for this outside of China for less than $5k.

It’s a bit mind-blowing to think that you could own a humanoid robot for less than $5k. It’s still a novelty at this point, and humanoid robots still can’t reliably complete your average day-to-day household chores but it’s looking increasingly likely that this is a solvable software problem, not a hardware problem now.

This means that it’s likely in the not too-distant future that people will be able to buy a genuinely helpful humanoid robot for less than $5k at today’s prices. If/when this takes off it’s going to have a massive impact on the world.

Forbes | Unitree

Web 4.0

AI Ethics News

Long Reads

Helen Toner - The term “AGI” is almost useless at this point
The New Yorker - Sam Altman May Control Our Future - Can He Be Trusted?

A week in Generative AI: Gemma, Microsoft & Emotions

Sean 🤓 — Sun, 05 Apr 2026 18:01:30 GMT

For your entertainment and delight this week we have three news stories that shine a light on a new battleground I think we have coming - speed and efficiency. First we have Google’s launch of Gemma 4, Microsoft launched three new frontier models, and then there is some great research from Anthropic on the ‘emotions’ of AI models.

In Web 4.0 news there’s a nice little update on ChatGPT Ads and a report on the stark reality digital news publishers are facing with the rise of AI platforms.

In Long Reads there’s a really good article from Stratechery celebrating 50 years of Apple which also has a great argument for why Apple is wining the AI war by sitting it out. Great argument, and a great read.

Google announces Gemma 4 open AI models

We’re now on the 4th generation of Google’s ‘open-weight’ Gemma models and the biggest change this time around is that they’re now released under a proper Apache 2.0 licence, the most common open-source licence there is. This is great for developers and businesses.

Gemma 4 is also a big update to the capabilities of this family of models, over the previous version that is 12 months old now. The models were built using all the same research and technology that were used to build the third generation Gemini models that have been hugely successful over the last 6 months.

Gemma 4 comes in the usual variants allowing them to be run locally, which dramatically reduces costs and increases speed. There are versions that can run on laptops and versions that can run on mobile devices.

YouTube | Google | Ars Technica

Microsoft takes on AI rivals with three new foundational models

Three big frontier AI models were released by Microsoft’s AI team this week - one for transcription, one for voice generation, and one for image generation.

A lot of Microsoft’s announcement focuses on speed and efficiency which I think will be an interesting battleground going forwards. Over the next couple of years I think we’re going to see a lot of progress here, especially on speed and we’ll look back at today’s AI like we view dial-up modem from the 90s.

TechCrunch | Microsoft

Anthropic Says That Claude Contains Its Own Kind of Emotions

This is a super interesting video. Put aside the idea of emotions for one second and what it’s actually about is identifying parts of the model that influence behaviour. Turn up/down certain parts and it’s more/less likely to cheat at a test.

The video concludes that we need to think of AI models as characters in their own story. They don’t have emotions, but there are aspects of its character that can influence it, just like our emotions influence us.

This could lead to better model tuning in the future for specific tasks and generally helping align models more successfully. Great work!

YouTube | Wired

Web 4.0

AI Ethics News

Long Reads

One Useful Thing - Claude Dispatch and the Power of Interfaces
Stratechery - Apple’s 50 Years of Integration
Business Insider - OpenAI needs to make money. Meet Fidji Simo, the ‘founder-mode’ executive charged with making it happen.

A week in Generative AI: Control, Pivots & Figure 03

Sean 🤓 — Sun, 29 Mar 2026 18:01:24 GMT

It was a quieter week this week in the mainstream AI news, but Anthropic quietly went about shipping a huge volume of developer-facing features and news continued to seep out about OpenAI’s pivot. Oh, and Figure 03 turned up at the White House.

In Ethics news, David Sacks has departed as the White House’s AI Czar, there’s more follow up to Anthropic’s battle with the Pentagon, and Wikipedia cracks down on the use of AI in article writing.

Good interview with Sundar Pichai in Long Reads too.

Claude can now control your Mac

Anthropic have had one helluva week, shipping Channels, Dispatch, Computer Use, Auto Mode, and iMessage support in the last 7 days. Much of this is developer-facing, but I think allowing your AI Assistant to control your computer from your phone is interesting more broadly.

Much of this is inspired by the approach OpenClaw took, primarily to give it access to more of the things that you usually do on your computer, including all your data, as well as a dedicated compute environment.

Anthropic’s approach is a bit different, but it also presupposes that you have a computer to connect to in the first place. Many consumers don’t have a personal computer at home and do everything they need to do on their phone, tablet, or work computer, and because of this I don’t think this will be a big hit with consumers, who would want something much more native and integrated into their mobile device.

Interesting ideas though!

Anthropic | Simon Willison | X

OpenAI Pivots

It’s no secret to anyone paying attention that OpenAI is going through a bit of an identity crisis at the moment. Or if it is supposed to be a secret, then it’s not being kept very well!

It’s been widely reported that OpenAI are pivoting to focus less on consumers and more on Enterprise. I think this comes from two places - one a recognition of the success that Anthropic has been having in this space, and two a desperate need to grow revenues faster. Part of this pivot is to drop all their ‘side quests’ which is why they just gave up on Sora (despite having recently inked a $1bn deal with Disney), erotic mode (🤮) and pulling back from Instant Checkout which they had been testing in the US for the last 6 months.

It’s the second reason that I think is the big driver here, but also the most confusing. Testing a paid ads model in ChatGPT is obviously a revenue driver, but so was Instant Checkout (they we taking a 2-4% cut of every sale) which is why I’m not sure why they’re pulling back from it.

It makes sense for OpenAI to stop all the ‘side quests’ they were on - it was getting a little ridiculous how many things they were trying to do at once and was starting to feel a bit like a ‘shotgun strategy’.

Focus is good. Revenue is good. But it still doesn’t feel (at least from the outside) that they’ve quite settled on what their strategy should be going forwards.

Reuters | TechCrunch | The Verge | TechCrunch

Melania Trump and AI powered robot named ‘Figure 03’ open White House summit

Yep…. this happened. Memes followed.

The Guardian | YouTube

AI Ethics News

Long Reads

Fast Company - Google CEO Sundar Pichai’s plan to make Gemini the only AI that matters
Artificial Ignorance - The Bots Are Reading Along

A week in Generative AI: Nvidia, GPT-5.4 & Bernie

Sean 🤓 — Sun, 22 Mar 2026 19:01:50 GMT

The big event of the week was Nvidia’s annual GTC conference that didn’t go by without raising a fair amount of controversy with the gaming community by using generative AI to make video games more photorealistic.

The big release of the week was, ironically, the mini and nano version of GPT-5.4 from OpenAI which is the first time we’ve seen smaller models from then since the original launch of GPT-5 last summer.

The big video of the week was Bernie speaking with Claude about AI and user privacy. It’s worth a watch as I think Claude comes across really well, but unsurprisingly it is telling Bernie all the things he’s likely to agree with!

The big Web 4.0 news is that Google is now using AI to replace headlines in search. It is not ok to change a publications headlines just to suite your needs, especially when doing so surreptitiously and without any notice or permission.

Nvidia’s DLSS 5 uses generative AI to boost photorealism in video games

This week, Nvidia had its annual GTC conference with the usual gang-buster 2 hour+ keynote from Jensen Huang. It covered everything from AI to gaming, to robotics, to Olaf from Frozen 🥶.

The most controversial announcement was gaming related where Nvidia introduced DLSS 5 which uses generative AI to make characters much more photorealistic. Needless to say that this didn’t go down well with gamers or developers and resulted in Jensen Huang just telling everyone that “they’re wrong”.

Many gamers view games as a form of art and by manipulating the graphics you’re changing the style and intent of the game designer. Yes, Nvidia say that developers and gamers are in control at all times, but to me this is another classic example of a generative AI solution in search of a problem.

Just because you can do something, doesn’t mean you should and given that many major game developers were blindsided by this announcement I see no evidence of Nvidia developing this feature in consultation with any developers or gamers.

I had started to hope that we were though the initial phase of generative AI product development where it’s just bolted onto lots (and lots!) of existing products and starting to see products that have been developed with generative AI from the ground-up. This is not one of those unfortunately.

Nvidia | YouTube | TechCrunch | Wired

OpenAI release GPT-5.4 mini and nano

We haven’t seen mini and nano model varieties from OpenAI since the initial launch of GPT-5 last summer and these update those, bringing more intelligence and the knowledge cut off up from May 2024 to August 2025.

We don’t really talk about the knowledge cutoff of models much anymore now they’re so good for searching and pulling back up-to-date information to inform their answers, but I think its still important, especially for developers who predominantly use these models ‘out-of-the-box’ and don’t necessarily give them access to live search.

Having said that, these new models are 3-4x more expensive to use that their GPT-5 counterparts, which is quite the price hike for developers. The models are definitely better, but whether they’re worth that price increase remains to be seen.

OpenAI | Simon Willison

Bernie speaks to Claude

Firstly, hats off to Bernie for leaning into AI, being open-minded but critical, and paying attention to what’s happening. He has a bit to learn on how AI works as there is sycophancy at play here and Claude is telling him what he wants to hear. But, he is making a much broader, more important point about user data and privacy.

Being European, these questions seem woefully out of date. We’ve had GDPR for 8 years now specifically to try and address all of these concerns. Whilst I wouldn’t point to GDPR as a great success story in its implementation at least the fundamental idea was right - to give consumers more transparency and control over their data. However, it’s also painfully clear that GDPR is not fit-for-purpose with AI platforms and needs a re-think. I’d say that GDPR is broadly being ignored when it comes to AI platforms and that’s not a good state of affairs for anyone.

I also think this is actually quite a good ad for Claude - it comes across really well and even makes Bernie laugh with it’s answer to why consumer information is being collected - it’s simple answer is ‘Money’.

YouTube | Gizmodo

Web 4.0

AI Ethics News

Long Reads

Stratechery - Agents Over Bubbles
Shubham Bose - The 49MB Web Page

A week in Generative AI: Meta, Moltbook & Helix

Sean 🤓 — Sun, 15 Mar 2026 19:02:18 GMT

A bit of a quieter week this one. Meta has reportedly delayed the rollout of its new AI model, despite a $1bn hiring spree last summer. They also acquired Moltbook 🤷‍♂️.

Apple celebrated its 50th birthday, which I couldn’t help but share and Figure AI showed off their Helix robot tidying an already incredibly tidy living room.

Good long read from Ethan Mollick this week too.

Meta Delays Rollout of New AI Model After Performance Concerns

Nearly 12 months ago, Meta released the 4th generation of their Llama series of open weight models that up until then had been the poster child of the open weight community. That release went very badly for them and they were caught gaming AI benchmarks.

This resulted in a complete pivot from Meta, led by Mark Zuckerberg, who acquired a 49% stake in Scale AI, poached their founder to lead their new Superintelligence Lab, and spent a reported $1bn on hiring as much AI talent as they could.

This pivot hasn’t resulted in any new model releases yet, and 12 months is a long time in AI. So even if you don't believe the rumours that Meta has delayed the rollout of its new models because of performance concerns, the fact that there have been no new models released since Llama 4 is telling. Alongside this, Meta are reportedly planning sweeping layoffs to offset their AI investments, so things aren’t looking great for them right now.

It’s a real shame as the Llama family of models were great and really led the field for the first couple of years of the current AI boom. If Meta had stuck to that strategy I think both they and the AI community would have been better for it.

New York Times | The Guardian | Gizmodo

Meta acquired Moltbook

Ok, buddy…

Very weird. Very very weird for Meta to acquire this… 🤷‍♂️

TechCrunch

Apple celebrates 50 years

Technically, not AI related but I couldn’t let Apple celebrate 50 years without a mention, especially as I wrote about how I Miss Steve Jobs this time last year.

A great post to celebrate from Tim Cook, quoting my favourite ad of all time.

Congrats Apple - here’s to the crazy ones!

Apple

Helix 02 Tidies the Living Room

Always good to see the progress Figure AI are making with their humanoid robots. I swear they’re deliberately trying to make the robot seem more human with how it carries the towel… also that is the cleanest living room I’ve seen in a while - not sure it really needed tidying TBH!

YouTube

Web 4.0

Skills are awesome. Originally developed by Anthropic it’s great to see these being widely adopted by OpenAI, just like MCP before it. They’re essentially re-usable instructions you can give to you AI assistant that it can automatically use without you having to prompt it specifically. They’re widely used in AI coding and now just starting to enter the enterprise. I think we’ll be seeing a lot more on them this year.

AI Ethics News

Long Reads

One Useful Thing - The Shape of the Thing

A week in Generative AI: Fallout, GPT-5.4 & Web Search

Sean 🤓 — Sun, 08 Mar 2026 19:01:35 GMT

This week has seen the continued fallout of Anthropic vs. ‘DoW’ and a big new model release from OpenAI - GPT-5.4. Both are a big deal - the Anthropic situation because of the future direction it could point to and GPT-5.4 because it feels like OpenAI’s GPT-5 generation models have finally hit their stride.

In Web 4.0 there’s a nice little update on how ChatGPT searches, and some interesting Ethics news for you all too.

Enjoy!

Fallout of Anthropic vs. ‘DoW’ continues

As predicted last week, the impact of Anthropic’s falling out with the Pentagon has spilled over into this week.

Firstly, Sam Altman has admitted that their rushed deal with the Pentagon looked ‘sloppy’ and OpenAI have had to admit that they can’t control the Pentagon’s use of AI and ultimately had to cave in on AI-powered surveillance. This has already led to one high profile resignation from OpenAI, their head of robotics Caitlin Kalinowski quit yesterday.

Unsurprisingly, Anthropic CEO Dario Amodei called OpenAI’s messaging around their military deal ‘straight up lies’ and the Pentagon officially labeled them a supply-chain risk, which Anthropic has said they will challenge in court. There’s also been a strong consumer reaction to all of this with ChatGPT uninstalls surging by 295% and Claude’s app installs surging.

If you’re interested about learning more or hearing more in-depth analysis and opinions I recommend the podcasts below:

OpenAI launches GPT-5.4 with Pro and Thinking versions

On Thursday OpenAI released their latest model, GPT-5.4 and I have to say that it is a very good model.

This feels like the GPT-5 model we’ve been waiting for and reminds me of how Apple’s operating system versions tend to work - there’s a lot of excitement about the new features for the major model releases, but it’s not until you get to the .3 or .4 releases 6 months later that you really start to see what they’re capable of.

And I think the same is true here. GPT-5.4 feels really good. The auto-switching between thinking modes is really dialled in, it’s fast when it needs to be, and takes its time when it needs to. It’s less error prone and the character of the model feels much nicer - no more of the sycophancy.

OpenAI touts better benchmark scores, lower token usage, and GPT-5.4 is a hybrid model that includes all the coding advancements they’re been building with the codex versions of the previous models. They’re also positioning it as the best model for professional work with improved tool and skill usage and a better foundational model for building agentic workflows.

I’m impressed.

TechCrunch | OpenAI

Web 4.0

How AI platforms search is one of the most important areas for marketers to understand as more and more people start using them in their day-to-day digital lives. That’s why whenever one of the major platforms makes some changes I pay attention.

Web Search on AI platforms is VERY different from how traditional search works. For example, consumers usually only perform 1-2 searches to answer their question and usually take the 5th result on average. AI platforms usually run 2-4 search queries in parallel and on average use the 12th result, which wouldn’t even appear on the first results page of a traditional search.

The reasons for these differences aren’t fully researched and understood yet, but I believe that AI platforms are very good at searching and finding results for the intent behind the prompt they’ve been given, whereas traditional search is driven more by the keywords in the search query. So AI platforms prioritise intent and relevance, whereas traditional search prioritises keywords and traffic volumes. Lots more to learn!

AI Ethics News

A week in Generative AI: War, Bananas & Kung Fu

Sean 🤓 — Sun, 01 Mar 2026 19:01:35 GMT

This week in AI has been dominated with the very messy, very public fallout between Anthropic and the US Government, which I’ve tried to outline and simplify for you all below. It’s inevitable that it will spill into next week, especially with OpenAI wading in and causing more confusion by claiming they’ve now done a deal with the ‘Department of War’ on the same terms Anthropic was sticking to, which the US government denies. We also had the release of Nano Banana 2 from Google and some impressive robotics Kung Fu that’s a fun watch.

In Web 4.0 news, OpenAI announced that ChatGPT now has 900m weekly users and in Ethics news, a thought experiment on Substack send the Stock Market Spiralling.

There are a few good long reads to check out this week too - The trap Anthropic built for itself (which relates to Anthropic vs. the US Government) and The billion-dollar infrastructure deals powering the AI boom.

Anthropic vs. the ‘Department of War’

Back in June 2024 Anthropic were the first frontier AI company to deploy their models in the US government’s classified network. Since then, Anthropic have always maintained two red lines as it related to the use of its technology for military operations: it will not allow its models to be used for mass domestic surveillance or to develop fully autonomous weapons that would operate without human involvement.

On Tuesday Pete Hegseth Demanded Anthropic Drop their AI Safety Guardrails, which was the culmination of an ongoing standoff between the company and the US Government. On Thursday Anthropic refused the Pentagon’s new terms. In a public statement Amodei said that whilst fully autonomous weapons may eventually “prove critical for our national defense,” that “today, frontier AI systems are simply not reliable enough to power fully autonomous weapons.”

On Friday, this all lead to employees at both Google and OpenAI supporting Anthropic’s stance in an open letter with over 700 signatures. Trump then ordered the US government to stop using Anthropic and Pete Hegseth designated them as a supply chain risk. Following this, Anthropic released a new statement on the comments from ‘Secretary of War’ Pete Hegseth.

In a confusing turn, Sam Altman then posted on Twitter that OpenAI had reached an agreement with the ‘Department of War’ to deploy their models that included protections addressing the same issues that were a flashpoint with Anthropic. The US government then contradicted this in their own post saying their agreement with OpenAI allowed for “all lawful use” of their models. OpenAI’s full announcement is here - Our agreement with the Department of War

What a mess. Anthropic stands to lose a contract worth up to $200 million and could be barred from working with other defence contractors and it’s unclear whether OpenAI’s agreement has the same safeguards or not. It’s great to see the AI community standing up for what’s right and coming together at the right time, but I’m not sure how aligned the leaders of the frontier AI companies are with their employees. They have very quickly been walking back many of the self-imposed safety measures they insisted were vital to the safe deployment of advanced AI systems, creating a mockery of any idea of self-regulation.

Incidentally, all of this has been great publicity for Anthropic, whose Claude app has rocketed to the top of the App Store rankings off the back of all the news stories this week.

Google launches Nano Banana 2 model with faster image generation

Google launched the original Nano Banana model in August and followed that up with a Pro version in November. Since the launch it’s been the ‘best of breed’ image model and supplanted OpenAI’s ChatGPT Image model that everyone got very excited about back in March.

Now they’re out with v2 which promises faster generation, more realistic images, and resolutions ranging from 512px to 4K, in different aspect ratios. Impressively Nano Banana 2 can maintain character consistency for up to five characters and fidelity of up to 14 objects in one workflow.

Nano Banana 2 becomes the default image model across all Gemini apps, so I’m sure lots of people are having a lot of fun with it already!

TechCrunch | Google

Unitree Spring Festival Gala Robots

A little bit of robotics fun to end the week - for the last couple of years Unitree have really been pushing the limits of how humanoid robots can move and what their ~~fleet~~ syndicate of robots can do, all in sync is incredibly impressive and fun to watch.

No idea if any of these robots can do anything practical or useful, but I think this is likely to become the ‘drone display’ of 2026.

YouTube

Web 4.0

AI Ethics News

Long Reads

TechCrunch - The trap Anthropic built for itself
The Verge - Why is AI so bad at reading PDFs?
TechCrunch - The billion-dollar infrastructure deals powering the AI boom

A week in Generative AI: Gemini, Sonnet & Lyria

Sean 🤓 — Sun, 22 Feb 2026 19:01:10 GMT

We’ve had a week of small little releases from the big players with Gemini Pro 3.1 and Sonnet 4.6 both introduced. Google also released Lyria 3, it’s music generating model, and Anthropic released an interesting research study into Agent Autonomy.

In Web 4.0 news, Perplexity has decided it doesn’t think advertising in AI platforms works after all, and in Ethics News we have a now recurring theme around how AI impacts mental health, and how companies are ‘AI-washing’ mass layoffs.

In Long Reads, Ethan Mollick has an update to his regular Guide to Which AI to Use in the Agentic Era and there’s a good article from Benedict Evans on How will OpenAI compete?

Google’s new Gemini Pro model has record benchmark scores — again

On Thursday Google bumped its Gemini Pro model up to version 3.1, and it looks like it’s a big step up in capability. Google is highlighting it’s reasoning capabilities and says that it is ‘a smarter, more capable baseline for complex problem-solving‘.

TechCrunch | Google

Anthropic releases Sonnet 4.6

Little bump in model number for Anthropic’s Sonnet (mid-tier) model that is the default for free and pro plan users. It doubles the context window to 1m tokens, which is a big help for coding tasks and Anthropic claim that early testers preferred it to their flagship Opus 4.5 model.

Anthropic are squarely aiming the release of Sonnet 4.6 at enterprise users, highlighting its improvements in Computer Use and automating ‘economically valuable office tasks‘.

It’s becoming increasingly clear that Anthropic is aiming itself at building out the infrastructure that enterprises will need to fully embrace agentic coworkers in the near future and their Sonnet 4.6 model is a big stepping stone in that direction.

TechCrunch | Anthropic

Google introduces Lyria 3

This latest music generating model from Google looks much more steerable and sophisticated than previous offerings and is one of the few features available in Gemini that ChatGPT doesn’t have. I wonder if we’ll see OpenAI add music generation to its platform this year?

TBH, I’m not sure how I feel about music generating models, in the same way I’m unsure about image and video generating models. I think they will be amazing tools and great for creative workers to quickly play with new ideas and mock things up but I desperately don’t want them to replace the skills, craft, and artistry that exists in our world. This is probably how people felt about paintings when cameras were invented and what we saw was an evolution of painting and the birth of a new art form in photography. I hope we see the same sort of thing play out this time too.

YouTube | Google

Anthropic’s Agent Autonomy study

Really interesting insights here from Anthropic on how people are using Claude Code. Unsurprisingly it’s mostly used for software engineering, but as time goes on and people have found other uses for it, we’re seeing new use cases pop up, which I suspect was the catalyst for Anthropic to release Cowork which is powered by Claude Code.

It’s also interested to see how long Claude Code works for before stopping (c. 40 minutes, at a 99.9% success rate). For anyone who’s seen me talk about METR’s research that measures AI agents ability to complete long tasks, this is a a great comparison point. I usually present the 80th percentile graph (at an 80% success rate - see below) which shows how AI agents have been improving exponentially over the past few years in this regard.

I think we’re soon going to be seeing agents that can successfully (99.9% reliable) complete software engineering tasks for over an hour and we’ll soon start to see that trickle out to other types of tasks like the ones Anthropic highlighted in their research.

AI News | Anthropic

Web 4.0

AI Ethics News

Long Reads

One Useful Thing - A Guide to Which AI to Use in the Agentic Era
Stratechery - Thin Is In
Benedict Evans - How will OpenAI compete?
New York Times - The A.I. Disruption Has Arrived, and It Sure Is Fun

A week in Generative AI: Exodus, Burnout & Deep Research

Sean 🤓 — Sun, 15 Feb 2026 19:00:49 GMT

This week’s newsletter is a bit more personal. Two AI-related stories caught my attention this week that directly challenge the productivity narrative we keep hearing. One shows people losing jobs to AI efficiencies. The other reveals what’s happening to those who keep their jobs: they’re not working less, they’re working more. As someone who uses AI heavily and has dealt with burnout, both hit uncomfortably close to home.

In more typical AI-related news though, there was just one thing that stood out, which was an update from OpenAI to Deep Research, a big feature I think is mostly slept on.

In Web 4.0 news, OpenAI started testing ads in ChatGPT, and Google previewed WebMCP, a new protocol for AI agent interactions. In Ethics news OpenAI disbanded its mission alignment team and Anthropic said it’ll try to keep its data centers from raising electricity costs.

I also highly recommend the Long Read from the NY Times - I Left My Job at OpenAI. Putting Ads on ChatGPT Was the Last Straw.

UK ad agencies undergo their biggest exodus of staff as AI threatens industry

It’s fair to say that the ad industry has had a tough few weeks. There have been big losses reported, as well as a sell-off fuelled by a wider sell-off in tech stocks following the announcement of Claude Cowork. There’s also been lots of agency consolidation and marketing budgets have flatlined.

All of this has been building for a while, and has resulted in advertising agency staff in the UK shrinking by 7% YoY with creative agencies taking a bigger hit with a 14% drop.

It’s very easy to blame this on AI, but the reality is that the advertising industry is going through huge structural changes at the moment. This is mostly driven by how the advertising ecosystem has been changing over the last 5-10 years with shifts towards performance marketing, the rise of influencers, and the dominance of the large digital platforms.

AI will absolutely drive a lot of change throughout the advertising industry, and sadly this is being looked at very one dimensionally at the moment. This is being driven by an incredibly competitive agency market that is driving down prices which agencies are hoping to recoup in the medium term by finding AI-driven efficiencies, even though these efficiencies are yet to be realised. AI technology is still maturing, the infrastructure needed to fully leverage it needs building out, and new skills and ways of working need to be established. The human cost of this transition is already being felt, particularly among younger workers and recent graduates who are finding it increasingly difficult to break into the industry.

My hope is that AI will open up a huge amount of opportunity alongside the change that it will drive. AI platforms are now used by c.1.5bn people every week worldwide and have matured into a viable marketing channel in their own right. Helping brands navigate this exciting and emerging space will require a lot of knowledge and expertise and advertising agencies should be well placed to provide this if they’re able to make the right investments and not just focus on cost cutting and efficiencies.

The Guardian

The first signs of burnout are coming from the people who embrace AI the most

While the previous article looked at job losses in the advertising industry, recent research published in Harvard Business Review reveals what’s already happening to those who retain their jobs. As someone who has suffered from burnout and is a big user of AI, this particularly resonated with me.

What this study found was that workers are adopting AI because it feels empowering, but this leads to taking on more tasks and doing more multitasking. The researchers describe this as a self-reinforcing cycle: AI accelerates tasks, which raises expectations for speed, which makes workers more reliant on AI, which expands the scope of what they attempt, which further increases workload. Workers found themselves taking on tasks that used to require specialists, prompting AI during lunch breaks, and juggling multiple AI-assisted workflows simultaneously, all of which felt productive in the moment but accumulated unsustainably.

This directly undermines the efficiency logic driving job cuts in the ad industry. When agencies reduce headcount expecting AI to fill the gap, they’re not creating sustainable efficiency, they’re redistributing work to fewer people who will work at higher intensity. The productivity gains are real in spreadsheet terms, but they’re achieved by extracting more from those who remain.

The danger is that the cognitive fatigue, burnout, and weakened decision-making that follows are lagging indicators - it will take time for the signs to appear. The question is whether organisations will recognise this dynamic before the initial productivity surge gives way to the inevitable crash. Because right now, most are doubling down on the efficiency narrative while the teams using AI every day are quietly drowning under an ever-expanding workload.

TechCrunch | HBR

OpenAI Updates Deep Research

OpenAI released some nice updates to Depp Research in ChatGPT this week allowing users to specify sources to search and to be able to update research with new sections and analysis after the initial run.

I think Deep Research is one of those ‘slept on’ capabilities in the AI world - hugely useful and powerful when you use it, but I don’t think many people do.

The Verge | YouTube

Web 4.0

AI Ethics News

Long Reads

NY Times - I Left My Job at OpenAI. Putting Ads on ChatGPT Was the Last Straw
Harvard Business Review - AI Doesn’t Reduce Work—It Intensifies It

A week in Generative AI: Opus, Frontier & Ads

Sean 🤓 — Sun, 08 Feb 2026 19:00:35 GMT

We had two big rival model releases this week - Claude Opus 4.6 from Anthropic and GPT-5.3 Codex from OpenAI. Both focused on coding. Both big deals for agentic AI. Anthropic also released their Super Bowl ad, taking a pop at OpenAI testing ads in ChatGPT, and Boston Dynamics released a fun Atlas video.

There’s a small number of Web 4.0 and Ethics News articles - Gemini has now surpassed 750m weekly users, very quickly closing the gap with ChatGPT. There’s also a report on the $100bn NVIDIA/OpenAI deal and the security challenges inherent in OpenClaw.

There are also a good couple of Long Reads to check out - Enjoy!

Anthropic releases Opus 4.6 with new ‘agent teams’

On Thursday Anthropic released the latest version of its flagship model, Opus, bumping it to 4.6, up from the 4.5 version which was only released in November last year.

The big focus of this update is on the model’s coding skills, especially as these skills are what are starting to unlock a lot of amazing agentic capabilities, including helping people with ‘everyday’ tasks in their Cowork application.

On the same day, OpenAI also introduced GPT-5.3 Codex. Again, the focus here is on coding capabilities, but OpenAI haven’t started applying this to helping people with ‘everyday’ tasks. I suspect we’ll see something big from OpenAI in this space in the next few months.

Mark Andreessen said in 2011 that "Software is eating the world". Now AI coding agents like Claude Code and Codex are eating software, and beginning to control the world by proxy.

Anthropic | TechCrunch | OpenAI

Anthropic says ‘Claude will remain ad-free,’ unlike ChatGPT

In Anthropic’s Super Bowl spot they have taken a pop at OpenAI for testing ads in ChatGPT. Sam Altman got exceptionally testy over this and the whole thing has highlighted that there is still a huge rivalry between the two companies.

For the last couple of years it has seemed like OpenAI were very focused on Google Deepmind as their main rival and Anthropic were happily focusing more on the enterprise market and catering specifically to developers with the coding capabilities of their models and Claude Code.

But as coding skills have become more central to agentic AI over the last 3 months the rivalry between OpenAI and Anthropic has really heated up, culminating in the joint releases of Opus 4.6 and GPT-5.3 Codex this week. It’s great to see the competition driving capabilities forward, and I hope they don’t get too distracted by taking pops at each other - there’s plenty of room for lots of different AI platforms in the future.

The Verge | TechCrunch

Atlas Airborne

The research version on Atlas has reached the end of the line, now that they’re heading into full production, so the researchers thought they’d really try and push the limits of what it could do. Pretty impressive!

YouTube

Web 4.0

AI Ethics News

Long Reads

Stratechery - Microsoft and Software Survival
Artificial Ignorance - The Codex App Has Upended My Daily Workflow
Latent Space - Experts Have World Models. LLMs Have Word Models

A week in Generative AI: OpenClaw, Genie & Helix 02

Sean 🤓 — Sun, 01 Feb 2026 19:01:09 GMT

This week the hype train around Molbot (now OpenClaw) really reached a crescendo, and I explain a little bit about what it is, and why I think it’s important, in this week’s newsletter below. Google also introduced their latest iteration of Project Genie, and Figure AI showed off Helix 02.

In Web 4.0 news, it’s been a long time coming but the UK’s CMA ruled that media groups should be allowed to opt out of Google AI Overviews without also opting out of Google Search and YouTube Overtook Reddit as the Go-To Citation Source for AI Platforms.

In Long Reads, Dario Amodei published his latest essay on The Adolescence of Technology, Ethan Mollick wrote about Management as AI superpower and Charlie Guo shared his thoughts on OpenClaw.

Moltbot (now OpenClaw), the AI agent that ‘actually does things,’ is tech’s new obsession

Ok, so we need to talk about Moltbot (previously Clawdbot, but now OpenClaw). There’s a chance not many of my readers have heard of it yet, but it has completely blown up in the last few weeks in tech circles.

In short, it’s an AI agent that anyone (with some tech know-how) can spin up and connect to lots of different services (calendar, email, notes, reminders, music, smart home appliances, you name it!). This allows the agent to do lots of different things for people and is semi-autonomous in that it can be proactive and be triggered by lots of different events.

The reason this has blown up is that people have seen what an unrestrictive, proactive personal assistant can do. But there are some drawbacks. Not only is it quite technical to set up (fixable) but there are also lots of obvious security and privacy risks with giving an AI agent unfettered access to some of your most personal services and content.

I think what this points to is two things. First, we’re seeing the pent up demand from Siri/Google Assistant still not being very capable play out. Secondly, we’re seeing the power of the agentic ecosystem that’s been built out - first with MCP and more recently with Skills. This means that it’s now relatively trivial to connect an AI agent to lots of different things. And if the connection you need hasn’t already been built then it’s relative quick and simple to build it yourself with agentic coding tools like Claude Code and Codex.

So what should we make of this? I think it’s very exciting and definitely points towards how useful AI agents could generally be for people. I think most consumers will see some of this when the next generation of Siri launches in the next couple of months. But I also think that it’s going to take some time for this technology to mature to the extent that it’s secure and private, but still very useful. I’m long on Apple being the ones to get this right, despite being (relatively) slow to the generative AI party.

And if you really want to blow your mind you should check out Moltbook, which Simon Willison says is the most interesting place on the internet right now. It’s essentially Reddit for people’s AI agents (not just OpenClaw). At the very least its a curiosity, and at its most extreme shows how weird an AGI “take-off” scenario might look if one happened for real.

This is definitely the weirdest, most interesting, and more exciting thing that’s happening in AI so far this year, so one to pay attention to.

Google Introduce Project Genie

I first wrote about Genie back in March 2024, and followed up writing about the third generation back in August last year. It’s been a fascinating project to watch and where it’s got to now is slightly mind blowing.

It doesn’t have an obvious use case yet beyond rapid prototyping of games, and you could see the potential here for VR, but even just watching the introductory video hints at what this could become.

My favourite example is probably when someone takes a picture of their cat which then allows them to explore the room in the photo from the cat’s perspective - bonkers!

The Verge | Project Genie | YouTube

Introducing Helix 02

NGL - I never thought I’d want to watch a video of something unpacking and repacking a dishwasher, but here we are in 2026 and there’s something strangely compelling about it.

This is a typical humanoid robotics demo video that we’ve seen a lot of before, but there are two important differences with this one. The first difference is that this is a single neural network controlling the robot, which should be more effective than previous approaches. The second is that the movement of the robot is starting to look more ‘human-like’ which I think is encouraging. If you want an example of what I mean, skip to 3:15 in the video where the robot uses its foot to close the dishwasher, something many of you have probably done hundreds of times without even thinking about it.

I think what we’re going to see over the coming years is humanoid robotic movement becoming more and more ‘human-like’, which will both make robots more effective at navigating our world, and simultaneously make us more comfortable around them. We’ll then just have to figure out what to do about the noise they make when they move!

Figure AI | YouTube

Web 4.0

AI Ethics News

Long Reads

Dario Amodei - The Adolescence of Technology
One Useful Thing - Management as AI superpower
Artificial Ignorance - Humans Welcome to Observe