Mathematics, agents and wallet security
OpenAI’s claimed mathematical results, practical agent workflows and a Coldcard entropy failure lead an edition where verification matters as much as capability.
Briefs
Frontier Research and Models
-
OpenAI’s reported mathematical advances — OpenAI says its internal Astra model produced ten results across mathematics and theoretical computer science, with arguments formalised in Lean. An assessment places the claims among major open problems, while commentary highlights the likely bottleneck: experts must still check, interpret and prioritise machine-generated proofs.
-
DeepSeek’s million-token inference result — A DeepSeek V4-Flash deployment repository reports 95.9 tokens per second at a one-million-token context across two DGX Spark systems. The result depends on this exact hardware and software configuration, rather than offering a general comparison.
-
Task cost, not token price — A reported cost comparison puts DeepSeek V4-Flash at $0.03 per task, versus $3.15 for Fable 5. Retries and extra turns can increase the final bill.
-
A tougher long-context test proposal — A proposed million-token literary prompt would test retrieval and reasoning across a long input, not just plausible short answers. It is an evaluation idea, not performance evidence.
Engineering Practice
-
Stacked diffs keep dependent work moving — Stacked diffs break dependent work into smaller pull requests that can be reviewed, tested and reverted separately. They can reduce waiting in large repositories, though teams need solid Git habits and suitable tooling.
-
Production agents need clearer harnesses — A practitioner discussion argues that harmful agent actions often expose vague specifications and weak safeguards. Production systems need clear limits, confirmation steps and strong protection around deletion or other irreversible actions.
Research Workflows
- Hermes makes literature reviews inspectable — Hermes stores review protocols, evidence records and intermediate decisions outside the chat. That makes a model-assisted literature review easier to inspect and resume, but human scientific judgement still decides whether its conclusions hold.
Cryptocurrency Security
- Coldcard entropy flaw and traced theft — A technical investigation links a Coldcard firmware change to weak wallet entropy. Galaxy Research’s analysis reports 1,158.81 BTC drained from 2,673 addresses into six collector addresses. Affected users should verify their wallets.
Tutorials & Demos
Google’s Antigravity agents build and repair an operating system
Google’s Antigravity demonstration shows 93 Gemini-powered sub-agents dividing an operating-system project into parallel jobs. The presentation says they made more than 15,000 model requests over 12 hours, creating scheduling, memory-management and file-system components. It then shows the agents repairing missing video and keyboard drivers after the initial build could not run Doom. The claimed API cost sits below $1,000, which makes the scale notable for early software work. Still, this is a vendor demonstration, not proof that the resulting system is complete, reliable or reproducible. Human review remains necessary when agents coordinate work across a large codebase.
Worth watching
- 26:09 — The challenge of building a working operating system with Antigravity and Gemini.
- 26:15 — The scale of the run: 93 sub-agents, 15,000 requests and 2.6 billion tokens.
- 27:38 — The generated system runs, then agents repair its missing drivers.
LangGraph workshop on controllable corrective RAG
This LangGraph workshop presents agents as explicit nodes, edges and shared state. It builds from a simple tool-using agent to corrective retrieval-augmented generation, where the system grades documents, rewrites weak queries and searches again when evidence fails. This structure gives developers visible control over how an agent moves between steps, rather than leaving it inside an open-ended loop. The workshop also tests models against several facts scattered through long contexts, where retrieval and reasoning become less reliable. These are teaching examples rather than independent production benchmarks, but the practical lesson is clear: long context does not remove the need for checks.
Worth watching