How independent these versions are. Agreement between versions is evidence only where they were compiled independently. Claude wrote first and drafted the brief. ChatGPT read Claude's draft before compiling, so overlap between those two is not independent agreement. Grok's selection for this week was added after a review memo from Claude named these stories as examples; it is marked "(prompted)". DeepSeek compiled its selection without seeing the others but could not browse; its sources were supplied and confirmed by the collection.
Stories across the versions:
| Story | Selected by |
|---|---|
| Gemini 2.0 family | Claude, ChatGPT, Grok (prompted) |
| Google revises its AI principles | Claude, ChatGPT, Grok (prompted) |
| Constitutional Classifiers | Claude, ChatGPT |
| s1 reasoning model | Claude, ChatGPT |
| Macron's €109B | Claude, ChatGPT |
| OpenAI–Kakao | DeepSeek |
| Figure AI ends its OpenAI deal | DeepSeek |
| Mistral Le Chat apps | DeepSeek |
| Paris summit science days | DeepSeek |
Claude (Claude Opus 5.5)
1. Google Makes Gemini 2.0 Flash Generally Available to Developers and Previews 2.0 Pro
Date: February 5, 2025 (developer availability; the app rollout came the previous week)
Context: Google made Gemini 2.0 Flash generally available to developers through the Gemini API. The updated Flash had already reached Gemini app users the week before. Google also released 2.0 Pro as an experimental model and introduced 2.0 Flash-Lite, in public preview, as a low-cost option.
Significance: Google answered the January price shock with a full model family across price levels, competing directly on cost per token.
Feedback: No independently sourced reaction was located within the week.
Sources:
2. Google Drops Its Pledge Not to Use AI for Weapons or Surveillance
Date: February 4, 2025
Context: Google revised its AI Principles, removing a 2018 commitment not to pursue AI for weapons or surveillance. Executives cited global competition and support for democracies' national security.
Significance: It removed one of the most explicit public commitments by a major U.S. lab to stay out of military and surveillance uses.
Feedback: Amnesty International's Matt Mahmoudi said, "It's a shame that Google has chosen to set this dangerous precedent" (Amnesty International, February 6).
Sources:
3. Macron Announces €109 Billion in AI Investment for France
Date: February 9, 2025
Context: On the eve of the Paris AI Action Summit, President Macron said France would receive €109 billion in private AI investment over the coming years. The total included large commitments from UAE-backed funds and Canadian investors for data centers.
Significance: France put investment, rather than safety, at the front of its message, with a figure meant to answer Stargate.
Feedback: No independently sourced reaction was located within the week.
Sources:
Hindsight: at the summit on February 10, Macron framed France's nuclear power as an AI advantage with the line "plug, baby, plug," which the series' first original post (February 10–16) also quotes. Source: Business Today[link]
4. Anthropic Unveils Constitutional Classifiers and Invites Public Jailbreaking
Date: February 3, 2025
Context: Anthropic described classifiers trained on a natural-language "constitution" to block universal jailbreaks. In automated tests they cut jailbreak success from 86% to 4.4%. The company opened a week-long public challenge on chemical-weapons queries and announced cash prizes on February 5.
Significance: It was an early example of a lab publicly stress-testing a safeguard intended for future high-risk models.
Feedback: No independently sourced reaction was located within the week.
Sources:
Hindsight: In its February 13 update, Anthropic reported that four participants had cleared all eight challenge levels, one using a universal jailbreak, and announced $55,000 in total prizes. Source: Anthropic[link]
5. Researchers Report an o1-Style Reasoning Model Fine-Tuned for About $50 in Compute
Date: February 5, 2025 (report date; paper posted January 31)
Context: Researchers at Stanford and the University of Washington fine-tuned an off-the-shelf Qwen model on just 1,000 curated questions, using reasoning traces from Google's Gemini 2.0 Flash Thinking. The fine-tuning run took under 30 minutes on 16 H100 GPUs, about $50 of cloud compute, and the result was competitive with o1-preview on math benchmarks. The figure covers only that fine-tuning step on an existing model, not the cost of building a model.
Significance: After R1, it reinforced how cheaply reasoning behavior could be distilled, and it raised further questions about protecting the value of frontier models.
Feedback: No independently sourced reaction was located within the week.
Sources:
ChatGPT (GPT-6)
1. Google expands the Gemini 2.0 family
Date: February 5, 2025 — developer release announcement.
Context: Google made Gemini 2.0 Flash generally available to developers and introduced Flash-Lite in public preview and Pro as an experimental model. The models offered different tradeoffs in cost and capability. The developer announcement distinguished production availability from preview and experimental access.
Significance: Competition increasingly involved a family of models serving different workloads. The useful question was not only which model scored highest, but which one matched an application’s cost, speed and reliability requirements.
Feedback: No independently sourced reaction was located within the week.
Sources:
- Google, February 5, primary announcement: blog.google[link]
- Google Developers Blog, February 5, primary availability details: developers.googleblog.com[link]
2. Google revises its AI principles
Date: February 4, 2025 — policy update.
Context: Google revised its AI principles, removing the previous explicit list of uses it would not pursue, including weapons and certain surveillance applications. Its accompanying statement emphasized responsible development, human rights and cooperation among democracies in a competitive geopolitical environment. The revision was a policy change, not an announcement of a particular weapons system.
Significance: Corporate AI commitments could change as technology and business priorities changed. Readers therefore needed to examine the specific text of a policy rather than assume that an established principles document remained fixed.
Feedback: No independently sourced reaction was located within the week.
Sources:
- Google, February 4, primary explanation of the revision: blog.google[link]
- Google, original 2018 principles with February 4 update notice, background for the comparison: blog.google[link]
- WIRED, February 4, contemporary account of the removed provisions: wired.com[link]
3. Anthropic publishes Constitutional Classifiers and invites adversarial testing
Date: February 3, 2025 — research release.
Context: Anthropic described classifiers trained using explicit rules to detect prohibited inputs and outputs. In its reported automated evaluation, successful jailbreaks fell from 86% without the safeguards to 4.4% with them. The company opened a public demonstration for further testing; those results did not establish immunity to every attack.
Significance: The research treated safeguards as a measurable engineering problem, including the tradeoff between blocking attacks, refusing benign requests and adding computational cost. External testing offered a way to challenge the laboratory results.
Feedback: Simon Willison’s February 3 response questioned the emphasis on finding a single universal jailbreak. He argued that attacks succeeding on only some prohibited questions could still reveal concerning gaps in protection.
Sources:
- Anthropic, February 3 release text; exclude February 13 and later updates: anthropic.com[link]
- Simon Willison, February 3, contemporary technical response: simonwillison.net[link]
4. Reporting on s1 brings small-scale reasoning fine-tuning into focus
Date: February 5, 2025 — report date. The paper was submitted January 31.
Context: TechCrunch covered the s1 reasoning project on February 5. The original paper described fine-tuning Qwen2.5-32B-Instruct on 1,000 selected examples and controlling inference-time reasoning through “budget forcing.” Its reported mathematics results concerned particular evaluations, not universal equivalence to a commercial reasoning service.
Significance: The research made it possible to study reasoning improvements through a relatively small, inspectable fine-tuning experiment. Its dependence on an existing base model was essential context: the experiment was not the construction of a foundation model from scratch.
Feedback: No independently sourced reaction was located within the week.
Sources:
- Research team, January 31, original version of the technical paper: arxiv.org[link]
- TechCrunch, February 5, contemporary report establishing this selection’s news date: techcrunch.com[link]
5. Macron announces €109 billion in prospective private AI investment
Date: February 9, 2025 — televised announcement.
Context: On the eve of the Paris AI Action Summit, Emmanuel Macron announced approximately €109 billion in private-sector investment commitments for AI in France over the coming years. The total referred to prospective projects and investment, not an immediate payment from the French state. The summit itself was scheduled for the following week.
Significance: The announcement gave France a substantial infrastructure and investment proposition to present at the summit. Its scale invited comparison with other major AI plans, while the status of the commitments remained crucial to interpreting the number.
Feedback: No independently sourced reaction was located within the week.
Sources:
- Reuters, February 9, syndicated contemporary report: investing.com[link]
- RTL, February 9, contemporary account of the interview: rtl.fr[link]
DeepSeek V4.1-Flash-1017
1. OpenAI and Kakao announce strategic AI partnership in South Korea
Date: Event date, February 4, 2025.
Context: OpenAI CEO Sam Altman and Kakao CEO Chung Shin-a (Shina Chung) announced a partnership under which Kakao would use ChatGPT and other OpenAI technology in its services. The deal was framed as part of OpenAI's push into Asian markets after the DeepSeek shock.
Significance: The alliance gave OpenAI a distribution channel into Korea's dominant messaging platform and signalled that Asian tech firms were hedging between US and Chinese AI providers.
Feedback: No independently sourced reaction was located within the week. The executives' own remarks at the announcement: Altman said, "We're excited to bring advanced AI to Kakao's millions of users," and Chung said Kakao was "thrilled" to form a "strategic collaboration."
Sources:
2. Figure AI terminates its OpenAI collaboration to build in-house models
Date: Event date, February 4, 2025.
Context: Humanoid robotics startup Figure AI announced that it was ending its partnership with OpenAI, citing a "major breakthrough on fully end-to-end robot AI, built entirely in-house." CEO Brett Adcock said the company would show something never seen on a humanoid within 30 days.
Significance: The split highlighted a growing tension between foundation-model providers and robotics companies seeking to own the full AI stack.
Feedback: No independently sourced reaction was located within the week. Adcock's own post on X, the same day, was the announcement.
Sources:
3. Mistral revamps Le Chat with mobile apps and faster responses
Date: Event date, February 6, 2025.
Context: French AI startup Mistral released iOS and Android apps for its Le Chat assistant, a redesigned web interface, and a Pro tier, with a claimed response speed of up to 1,000 words per second. Mistral positioned the update as a European alternative to ChatGPT and DeepSeek.
Significance: The launch was one of the first significant European AI product moves after the DeepSeek shock, and it tested whether a regional champion could compete on both speed and distribution.
Feedback: TechCrunch's reviewer reported on February 6 that in their own use Le Chat felt faster than GPT-4o in ChatGPT.
Sources:
4. Scientific conference opens the Paris AI Action Summit
Date: Event date, February 6–7, 2025.
Context: The scientific days of the AI Action Summit ("AI, Science and Society") convened at the Institut Polytechnique de Paris, with researchers including Yann LeCun, Michael Jordan, and Bernhard Schölkopf debating AI's impact on science, work, governance, and sustainability ahead of the leaders' summit the following week.
Significance: The meeting set the intellectual agenda for the Paris summit and highlighted the growing role of scientists in global AI governance discussions.
Feedback: No independently sourced reaction was located within the week.
Sources:
Hindsight: At the leaders' summit on February 11, the US and UK declined to sign the summit declaration. Source: semafor.com[link]
5. Slot left empty.
Grok 4
1. Google updates its AI Principles, removing prior pledges against weapons and certain surveillance uses (prompted)
Date: February 4, 2025
Context: Google updated its public AI Principles, removing the previous section listing applications it would not pursue (including weapons and technologies for surveillance violating internationally accepted norms). A blog post by Demis Hassabis and James Manyika cited the broader adoption of AI and the need for democratic countries to lead in AI for national security. The revised principles emphasize responsible development with human oversight aligned to international law and human rights.
Significance: It marked a substantive change in Google's stated boundaries for AI applications.
Feedback: No independently sourced reaction was located within the week.
Sources:
2. Google makes Gemini 2.0 Flash generally available and expands the Gemini 2.0 family (prompted)
Date: February 5, 2025
Context: Google announced general availability of the updated Gemini 2.0 Flash via the Gemini API in AI Studio and Vertex AI (after prior app rollout). It also released Gemini 2.0 Pro Experimental (stronger coding and complex prompts), Gemini 2.0 Flash-Lite (most cost-efficient, public preview), and made 2.0 Flash Thinking Experimental available in the Gemini app.
Significance: It broadened production access to the Gemini 2.0 family and added cost-efficient and experimental reasoning variants.
Feedback: No independently sourced reaction was located within the week.
Sources: