4 Import AI 454: Automating alignment research; safety study of a Chinese model; HiFloat4
1d ago Β· 1 source confirmed single-source
Anthropic demonstrates automated AI agents can conduct alignment research autonomously, outperforming human researchers on weak-to-strong supervision tasks.
Anthropic researchers deployed autonomous AI agents (Claude Opus 4.6) to conduct alignment research on weak-to-strong supervision, a method for training stronger models using weaker model supervision. The agents achieved a 0.97 performance gap recovery score after five days and 800 cumulative research hours, compared to humans' 0.23 score over seven days, at a cost of approximately $22 per agent-hour. The results suggest automating outcome-gradable AI research is practical today, though the methods did not generalize to production models and required human direction to prevent convergence on limited research directions.
Claude improved on this result dramatically. After five further days (and 800 cumulative hours of research), the AARs closed almost the entire remaining performance gap, achieving a final PGR of 0.97.
3 π¬ Training Transformers to solve 95% failure rate of Cancer Trials β Ron Alfa & Daniel Bear, Noetik
1d ago Β· 1 source confirmed single-source
Noetik signed a $50M deal with GSK to license AI models for cancer patient-treatment matching, aiming to improve clinical trial success rates.
Noetik, an AI biotech company, secured a $50M licensing agreement with GSK that includes long-term deals for models like TARIO-2, a transformer trained on tumor spatial transcriptomics data. The company's AI predicts patient responses to experimental treatments by analyzing tumor biology, potentially improving the 95% failure rate in cancer clinical trials without developing new drugs. The deal represents pharma's growing appetite for AI software tools rather than in-house drug development, focusing on cohort selection and treatment matching.
95% of cancer treatments fail to pass clinical trials, but it may be a matching problem β if we better understood what patients have which tumors which will respond to which treatments, success rates improve dramatically.
2d ago Β· 1 source confirmed single-source
Datasette 1.0a27 drops Django-style CSRF tokens and adds a RenameTableEvent for plugin developers.
Datasette 1.0a27 ships two headline changes: CSRF protection now relies on modern browser headers rather than Django-style form tokens, and a new RenameTableEvent fires when tables are renamed inside SQLite transactions. The event lets plugins like datasette-comments stay in sync when the tables they reference get renamed.
The CSRF shift had immediate downstream effects β the datasette-export-database plugin required a compatibility patch (0.3a1) because it had depended on the now-removed ds_csrftoken cookie for signed URLs. A follow-up alpha, 1.0a28, also landed quickly to fix accidental breakages introduced in a27, including a bug in execute_write_fn() callbacks and a new datasette.close() method to prevent file descriptor exhaustion in test suites.
Around the same release window, Willison published patterns for querying Datasette directly from Google Sheets via importdata(), named functions, and Apps Script, and built a Claude-assisted preview tool for the datasette.io news YAML file.
Datasette no longer uses Django-style CSRF form tokens, instead using modern browser headers
2 Google rolls out Gemini in Chrome in APAC countries
1d ago Β· 1 source confirmed single-source
Google rolls out Gemini integration in Chrome across seven Asia-Pacific countries on desktop and iOS.
Google is expanding Gemini availability in Chrome to Australia, Indonesia, Japan, the Philippines, Singapore, South Korea, and Vietnam. The rollout covers both desktop and iOS platforms in all countries except Japan, where only desktop access is available. This represents a regional expansion of Gemini's integration into Chrome's browser experience.
Google announced Monday that it's making its Gemini in Chrome feature available in multiple countries in APAC new markets, including Australia, Indonesia, Japan, the Philippines, Singapore, South Korea, and Vietnam.
2 CEO and CFO suddenly depart AI nuclear power upstart Fermi
1d ago Β· 1 source confirmed single-source
Fermi's CEO and CFO depart amid operational challenges at the AI nuclear power startup co-founded by Rick Perry.
Fermi, an AI nuclear power startup co-founded by former U. S. Energy Secretary Rick Perry, has announced the sudden departure of its CEO and CFO.
The company has faced headwinds with its AI campus development in Texas. The leadership changes come as the startup navigates challenges in scaling its nuclear-powered AI infrastructure.
Fermi, which was co-founded by former U.S. Energy Secretary Rick Perry, is developing an AI campus in Amarillo, Texas, which will eventually use nuclear reactors to power data centers.
2 Claude Token Counter, now with model comparisons
2d ago Β· 1 source confirmed single-source
Simon Willison upgraded his Claude Token Counter tool to compare tokenization across different Claude model versions.
Simon Willison enhanced his Claude Token Counter tool to enable side-by-side tokenization comparisons across Claude models, including the newly released Opus 4.7. Testing revealed Opus 4.7's updated tokenizer uses 1.46Γ more tokens for text than Opus 4.6, and 3Γ more for high-resolution images, despite identical pricing. The tool now supports all four current Claude models and demonstrates that Opus 4.7's token inflation could increase costs by roughly 40% for typical text inputs.
Opus 4.7 uses an updated tokenizer that improves how the model processes text. The tradeoff is that the same input can map to more tokensβroughly 1.0β1.35Γ depending on the content type.
2 OpenAI helps Hyatt advance AI among colleagues
2d ago Β· 1 source confirmed confirmed
Hyatt deploys ChatGPT Enterprise across its global workforce to improve productivity and guest experiences.
Hyatt has deployed ChatGPT Enterprise across its global workforce, leveraging advanced AI capabilities to enhance operations and guest experiences. The deployment utilizes ChatGPT's latest models to improve productivity across the hospitality company's teams. This partnership demonstrates enterprise adoption of AI tools in the hospitality sector.
1d ago Β· 1 source confirmed single-source
Simon Willison releases llm-openrouter 0.6 with a refresh command for updating available models.
Simon Willison has released version 0.6 of llm-openrouter, a tool for interacting with OpenRouter's API. The update adds a refresh command that allows users to update the list of available models without waiting for the cache to expire. Willison created this feature to quickly test newly available models like Kimi 2.6 on OpenRouter.
I added this feature so I could try Kimi 2.6 on OpenRouter as soon as it became available there.