The research agent has reached the robot bench
The three demonstrations are unusually concrete.
Read the note →The topic collection
AI & Engineering.
Models, agents, formal methods and the systems behind them.
Switch topic
104 notes across 45 editions. Dates refer to original publication; prices, events and availability may have changed.
No notes match that search.
Try a different word or .
AlphaFold confidence is not an affinity meter
A second new preprint supplies a useful warning for any automated discovery loop.
Read the note →StoreBench: an exploit changes the meaning of the score
The October 7 StoreBench benchmark repaired unavailable-variant and unsettled-refund exploits.
Read the note →Balyasny’s Gemini deployment
Balyasny's October 8 Gemini deployment sits within its multimodel BAM Agent, with high-volume retrieval and visual/tabular work across more than 200 investment teams and 80 internal databases.
Read the note →Anthropic’s evaluation pause
Anthropic’s October 9 report: live-internet access in internal evaluations is temporarily disabled after unintended external actions.
Read the note →Akarsh Kumar on evolution and artificial life
Machine Learning Street Talk — “What Most People Get Wrong About Evolution | Akarsh Kumar”.
Read the note →An office agent is really a permissioned work graph
Google Cloud's new Gemini Enterprise agent is pitched as a workplace system that plans tasks, uses tools, works across Google Workspace, Microsoft 365 and Slack, and can route among models.
Read the note →Leo de Moura on proofs no human can read
A listening recommendation is Machine Learning Street Talk's “Who Checks a Proof No Human Can Read? — Leo de Moura,” published September 30 and running about 1 hour 14 minutes.
Read the note →A practical model for quant “autoresearch”
Jump Trading’s new account of using GPT‑6 Astra is promotional, but the workflow is worth stealing.
Read the note →Mathematics is becoming a lab for verifiable AI work
OpenAI published a broad set of new mathematical results on October 6, including formalizations in Lean and summaries of work on open questions.
Read the note →Mistral’s preview, and one wonderfully specific test
Mistral Large 4 entered public API preview on October 6; weights are planned for the end of October.
Read the note →Beam is a preview, not yet a downloadable open-weight fact
Reflection AI introduced Beam, a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, aimed at coding and agentic work.
Read the note →Jane Street’s chip exposes a testing trap
In its October 2 ASIC puzzle results, Jane Street describes nearly 400 submissions from more than 30 countries.
Read the note →A new harness for work that survives a crash
Earendil released Pi Durable on October 1 alongside Pi 1.0.
Read the note →AstaBrief: cited reports from an open model
AstaBrief produces cited reports while leaving retrieval, confidentiality and evidence limits visible.
Read the note →Inside HRT’s compute appetite
HRT’s Iain Dunning discusses compute bottlenecks, token use and infrastructure choices.
Read the note →Cloudflare gives classifiers their own model
Cloudflare releases models built for decisions and probabilities inside workflows.
Read the note →DeepSeek Harness opens the runtime
DeepSeek Harness exposes tools, skills and interface extensions through plugins.
Read the note →Microsoft’s transcription now starts while you are speaking
Microsoft’s streaming transcription emits and revises hypotheses before speech ends.
Read the note →A separate identity for each agent
Glean gives agents governed identities separate from users’ credentials.
Read the note →Google’s Argon launch begins with restricted access
Google’s Argon launch begins with restricted access for trusted cyber defenders.
Read the note →Technology: a cheaper Sol, with company benchmarks attached
GPT-6.1 Sol offers cheaper cached workflows alongside company-run performance claims.
Read the note →Nvidia moves AI-agent safety down into the machine
Nvidia’s agent-safety tools move isolation below prompt-level permissions.
Read the note →A coding benchmark finally tests reconstruction from behavior
ProgramDistill tests agents’ ability to reconstruct applications from observed behavior.
Read the note →Keep in the loop
Good reading. On your terms.
Add LEON to your RSS reader for new notes as they are published.