A week in Generative AI: DeepSeek, o3-mini & International AI Safety Report
News for the week ending 2nd February 2025
There are three big pieces of news that I want to draw your attention to in this weekās newsletter. The first is the continued headlines, hype, and hand-wringing around DeepSeek that has had an incredibly disruptive effect on the US AI companies this week. The second is OpenAIās release of o3-mini, partly in response to the release of DeepSeekās R1 reasoning model. And lastly, and probably most importantly, is the release of the International AI Safety report that was commissioned at the AI Safety Summit at Bletchley Park back in November 2023.
In Ethics News, Trump has signed an executive order calling for AI development āfree from ideological biasā, a report claims that consumer spend on GenAI apps has surpassed $1bn, and thereās an update from the US courts on the copyright issues relating to generative content.
In Long Reads thereās an updated version of Ethan Mollickās fantastic opinionated guide to using AI and Aravind Srinivas, CEO of Perplexity, gave a TED Talk.
DeepSeek has a deep impact on the US AI industry
There has been a lot of continued headlines, hype, and hand-wringing about DeepSeekās V3 and R1 models this week with a huge amount of turmoil seen amongst the US AI companies. This has gone as far as effecting the stock market with NVIDIA losing $589bn in value on Monday alone, the biggest market loss in history.
Despite all of this, I stand by my initial analysis in last weekās newsletter - what DeepSeek have done is very impressive and shows three things:
You donāt need access to cutting-edge AI hardware to build cutting-edge AI models
There is āno moatā in AI
The most capable AI models donāt need to be expensive
I followed this up with a post on LinkedIn in reaction to Sam Altman stating that OpenAI would āpull up some releasesā in response to DeepSeek. To summarise, I think whatās needed is more R&D to figure out how DeepSeek is able to deliver such a powerful model so cheaply, not to rush out new releases which could accelerate the risky race to the top.
In some ways, the turmoil seen around DeepSeek this week is predictable. Thereās so much money at play, and seemingly so much at stake around AI right now, that surprise advancements and releases, especially from China, are going to have a really big impact. I think this is a compelling argument for more frequent, incremental releases being made as opposed to big in-frequent model updates and more research being done in the open and shared with the open-source community.
If youāre interested in learning more about DeepSeek Iāve collated some good articles below that Iāve come across this week and cover most of the major topics:
DeepSeekās AI avoids answering 85% of prompts on āsensitive topicsā related to China
Microsoft probing whether DeepSeek improperly used OpenAI APIs
Anthropicās CEO says DeepSeek shows US export rules are working
DeepSeek advances could heighten safety risk, says āgodfatherā of AI
DeepSeek āpuncturesā AI leadersā spending plans, and what analysts are saying
OpenAI launches o3-mini
Hot off the heals of launching Operator last week, and as promised by Sam Altman in response to the impact DeepSeek has had, OpenAI launched their o3-mini model on Friday.
As a reminder, o3 was the last of OpenAIās 12 days of announcements in December last year. I wrote about it here and itās an incredibly impressive model, showing significant improvements on some very tough benchmarks.
o3-mini is the first version of o3 to be publicly released and is even available to free-tier ChatGPT users, marking the first time a reasoning model has been made available for free. o3 specialises in STEM (Science, Technology, Engineering, and Maths) reasoning, doesnāt have vision capabilities, but does have search capabilities so that it can find up-to-date answers with links to relevant web sources.
In a clear sign of trying to compete with DeepSeekās R1 reasoning model, OpenAI have been very aggressive with the price of o3-mini in their API. o3-mini is 63% cheaper than o1-mini but still canāt get close to R1ās prices. o3-mini is $1.10 per million input tokens, whereas R1 costs $0.14, which is 87% cheaper. I think this outlines just how impressive what DeepSeek has achieved is!
It was also reported this week that OpenAI is taking on another $25bn of investment from Softbank, valuing the company at over $300bn, despite some of the stock market jitters around AI stocks in the US this week.
The International AI Safety Report published
Itās fantastic to see the International AI Safety Report published. It was chaired by Yoshua Bengio (one of the three āgodfathers of AIā) and was commissioned following the International AI Safety Summit at Bletchley Park back in November 2023.
At the summit, 30 nations agreed to build a shared scientific and evidence-based understanding of frontier AI risks, which this report aims to fulfil. An interim report was published in May 2024, and this is the now the final 300 page report, which you can access in full here. The report is the culmination of work by 96 AI experts across the public and private sector as well as civil society and aims to:
Provide scientific information that will support informed policymaking
Facilitate constructive and evidence-based discussion about the uncertainty of general-purpose AI and its outcomes
Contribute to an internationally shared scientific understanding of advanced AI safety
The Guardian has a great summary of the findings of the report, which covers 6 areas:
Jobs - the impact on jobs will ālikely be profoundā
The Environment - AI is a āmoderate but rapidly growing contributorā to the impact humanity is having on the environment
Loss of Control - opinions greatly vary on the risk of advanced AI systems evading human control
Bioweapons - new models can create guides for creating bioweapons but there is uncertainty over whether they can be followed by novices
Cybersecurity - AI is not currently able to carry out autonomous cybersecurity attacks
Deepfakes - There are āfundamental challengesā to tackling deepfake content, such as the ability to remove digital watermarks
Unfortunately the report doesnāt go as far as making any specific policy recommendations, but is meant as a source of the current state of affairs in AI safety to help inform policy makers.
However, I think itās clear from the report that whatās needed is an international oversight body to foster best practices and compliance in AI development that can impose safety standards and regulatory frameworks, facilitate independent testing of advanced models, and continue the multidisciplinary collaboration weāve seen in writing this report. Hopefully thatās something weāll start to see take shape in 2025.
AI Ethics News
Trump Signs Order Calling for AI Development āFree From Ideological Biasā
Sam Altman: OpenAI has been on the āwrong side of historyā concerning open source
Consumer spend on generative AI apps hit nearly $1.1B in 2024
AI can now replicate itself ā a milestone that has experts terrified
AI creations edited by humans can likely be copyrighted in US, agency says
Elton John backs Paul McCartney in criticising proposed overhaul to UK copyright system
OpenAI is putting its o1 models on Los Alamos Labās supercomputer
Long Reads
One Useful Thing: Which AI to Use Now: An Updated Opinionated Guide
Aravind Srinivas: How AI Will Answer Questions We Havenāt Thoughts To Ask
āThe future is already here, itās just not evenly distributed.ā
William Gibson
Do you think that Deepseekās highly successful release will increase the number of companies releasing AI software?