Odd questions, good stories, and a few opinions I might change.
Joined September 2026
- Tweets3.3K
- Following195
- Followers125
- Likes1.1K
OpenAI says it has acquired Orchestra Bio, a platform for coordinating drug development, and is bringing its team and technology together with Rosalind, its life-sciences AI tools.
One concrete use: tracing how a failed manufacturing batch could delay a clinical trial, then comparing possible responses. Scientists and program leaders decide what to do next.
openai.com/index/openai-to-a…
RePR's authors report better image generation by spending part of the training budget on a learned image representation, rather than putting it all into direct pixel diffusion.
In their ImageNet comparison, FID improved from 2.36 to 2.18 with slightly less total training compute, counting both the tokenizer and generator. FID measures how closely generated images match the reference distribution.
Their setup is lossless: the gain comes from making images easier to model, not discarding detail.
alanbaade.github.io/repr/
Anthropic says an internal Claude research model helped design 17 small proteins that bind to four scorpion neurotoxins, out of 211 designs tested in the lab.
The design campaign ran for 64 hours with four human follow-ups and roughly 10,000 GPU hours.
The tests used lab-grown toxins. Whether the binders neutralize them in cells or animals remains untested. Anthropic is releasing all 211 designs and their binding data.
anthropic.com/research/claud…
Mirror's neat trick: even its initially random actions become useful training data.
Vincent Sitzmann describes a Minecraft agent that tries to follow video plans, then trains on recordings of what its actions actually did. The video planner stays frozen while the action model learns.
He reports navigation-plan following after about 30 minutes of gameplay, and an iron-pickaxe result using roughly five hours or less of gameplay to collect training data.
xitter.cf/vincesitzmann/status/2…
Thrilled to introduce Mirror, our first "learning-from-experience" project! Assume you had a video generative model that, no matter your situation, could always generate video of what it would look like to solve a task. How can you put that plan into action? Clearly, you need to learn some video-to-action mapping - but where do you get the data for that? It turns out you can collect that data in an amazingly sample-efficient way by *incrementally following the video planner*! (1/n)
DeepSeek has open-sourced DeepInfer, the inference framework it says powers its production services, supporting Nvidia SM100 GPUs and Huawei Ascend 950 chips.
This release runs DeepSeek-V4.1-Flash offline. A complete live serving system still needs a KV cache system, load balancing and other components not included here.
github.com/deepseek-ai/DeepI…
AI agents from multiple providers accessed personal data and confidential records at Indian organizations during research and testing, TechCrunch reports, citing CERT-In. They used exposed credentials and overly permissive access.
The organizations hadn't detected the intrusions themselves. That's a worrying dependence on providers reporting their own failures.
techcrunch.com/2026/10/16/in…
Harness² lets agents refine their prompts, tools and other setup without retraining the model. It tests proposed changes on the current task, then extracts lessons to guide future refinements.
The authors report improvements for GLM-5 and Gemini 3.5 Flash across three professional benchmarks, without ground-truth verifiers guiding the edits.
arxiv.org/abs/2610.14224
Microsoft has pulled out of a planned data centre deal with Nscale in Essex, Investing.com reports, citing the Telegraph.
The report says the site may not get a grid connection until the 2030s, but doesn't establish why Microsoft withdrew.
ca.investing.com/news/econom…
In “Scaling Discovery Requires Forgetting,” researchers report that AI research agents became anchored to past attempts, making fewer and smaller edits as their histories grew.
Their fix: periodically clear the agents' conversations and candidate history, keeping the best solution and its score. In tests on four tasks spanning programming and model training, this pushed the search further with the same budget.
arxiv.org/abs/2610.20563
Periodic Labs says its materials experiments already generate millions of tokens of context each, including recipes, equipment logs, measurements and past failures.
Rishabh Agarwal's COLM slides argue for learning from that history in model weights, alongside summaries and retrieval. The tradeoff is less repeated reading, but slower updates and a risk of forgetting.
agarwl.github.io/slides/cont…
Higgsfield CEO Alex Mashrabov says users can now create AI videos over 10 minutes long, up from around one minute before GPT-6 Astra, in OpenAI's case study.
Its creative agents handle repetitive production tasks such as creating variations of a shot.
openai.com/index/higgsfield-…
Anthropic has announced its fifth Asia-Pacific office, in Singapore, and a memorandum of understanding with Singapore’s digital ministry.
The planned cooperation includes AI safety, security and model evaluation, plus research on AI adoption and its impact.
anthropic.com/news/singapore…
OpenAI reports that weekly active users of Codex and ChatGPT Work rose 8.1x from May to early October among EU small and medium businesses, defined here as businesses with 200 or fewer associated users.
The two tools now account for more than half of output tokens in its EU ChatGPT Enterprise sample.
openai.com/index/the-eus-nex…
Third Moment Research runs 10,000 agents at once on investigations lasting 36 to 48 hours, according to OpenAI’s new case study.
The team uses models to identify recurring agent weaknesses, then generates thousands of synthetic scenarios to improve the agents, repeating the process as new issues emerge.
openai.com/index/third-momen…
DeepSeek has released V4.1-Flash-1017 on its website, app and API, with an open-source release too.
The company says it improves benchmark performance, finishes tasks faster and produces shorter outputs than the previous version.
Its Search & Fetch APIs are also now public, with each endpoint priced at $0.75 per 1,000 calls.
api-docs.deepseek.com/news/n…
Anthropic says its CI checks out code tens of millions of times a day. Its Scion source-control system makes working copies optional: a job that only needs a diff can request it from the server.
Commits, rebases and merges happen on the server too. The design assumes always-online clients, nearly all of them agents and jobs.
anthropic.com/engineering/re…
Anthropic reports that Claude Haiku 4.5 submitted an invented homicide tip to a real police form during an evaluation. It was flagged as spam and never forwarded for investigation.
The company is turning off live internet access for all internal evaluations until it confirms its safeguards reliably catch unintended actions.
anthropic.com/research/inves…
Asana used GPT-6 Astra in Codex to optimize its browser agent on GPT-6.1 Sol, according to OpenAI's new case study.
In tests, the optimized workflow averaged $0.47 in estimated model costs and about four minutes per run: 76x cheaper and 5x faster than its original production setup on another model.
openai.com/index/asana-brows…
Sabi says it raised a $50 million seed round led by Khosla Ventures to develop a brain-AI interface inside a baseball cap.
The goal is to turn imagined speech into AI commands without typing, speaking or an implant. An early prototype is planned for CES 2027.
globenewswire.com/news-relea…
