We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Safety and control
A prototype AI worm uses stolen compute to adapt and spread
Researchers from the University of Toronto, Vector Institute, University of Cambridge and ServiceNow describe a worm that runs an open-weight LLM on compromised machines, generates attack strategies for each target and propagates across Linux, Windows and IoT devices by exploiting common corporate-network vulnerabilities. The paper says stolen compute gives the attacker zero marginal cost per infection and makes safeguards tied to commercial AI APIs structurally irrelevant.
Import AI reports roughly 80% success in vulnerability detection, 53% in exploitation and 88% in self-replication—about 37% for a complete attack—using a harness with separate planning, judging, action, summary and progress nodes. It remains a proof of concept rather than a live-incident report, but the paper’s decentralized swarm design means the important shift is from fixed exploit code toward an agent that can adapt and retry across targets.
Frontier-lab employees ask for pacing tools
A statement signed by 1,346 employees of frontier AI companies says leading labs may be close to automating AI research and asks the US government to support an international effort to develop technical and governance tools to “deliberately pace” frontier automated AI development. Its rationale is a coordination problem: companies and countries face pressure not to slow unilaterally, while the world lacks tools to buy time for security and oversight.
Import AI says the signatories include chief scientists and cofounders from OpenAI, Anthropic, Google DeepMind, Meta and Safe Superintelligence, among others. The proposal is framed as a way to create room for safeguards, not as a unilateral halt to research.
Capability is advancing unevenly
OpenAI puts Astra’s ten mathematics results into an audit workflow
OpenAI’s official account says an internal version of Astra resolved or substantially advanced ten long-standing problems across geometry, coding theory, circuit complexity, group theory, operator algebras, quantum complexity, lattice cryptography and extremal combinatorics. The company estimates roughly $2,000 in token cost at Sol API rates; humans prepared the manuscripts, Astra formalized each argument in Lean, and OpenAI is releasing the manuscripts, certificates and reasoning walkthroughs for examination.
OpenAI says it takes responsibility for correctness and that the mathematical arguments themselves were generated by its system. That makes expert scrutiny the next meaningful test: the release supplies material for mathematicians to examine, but the announcement is not a substitute for wider validation.
Shadow evaluation finds the open-ended research bottleneck
A separate arXiv study introduces “shadow evaluations,” in which an agent tackles the central question of an unpublished paper and the original authors grade the result. In two unpublished NeurIPS 2026 case studies, frontier agents received six days and thousands of dollars of compute, completed the engineering without human help, but made no substantial progress on the research questions; both papers were rejected, with recurring failures in research judgment, creativity, backtracking, resource awareness and instruction-following.
The contrast is useful rather than contradictory: human-defined, verification-friendly problems can now produce striking results, while choosing worthwhile questions and abandoning bad approaches remain unresolved. The study calls this early evidence from only two case studies, so it narrows claims about research automation rather than settling them.
Products and economics
DeepSeek’s flash update makes post-training the battleground
A Two Minute Papers review reports that DeepSeek’s updated flash model more than doubled many benchmark results, improved one result sevenfold, and beat both its previous flash version and a pro model roughly five times larger. The review says the architecture and size did not change; the gain came from post-training that taught the model when to plan, check its work and recover from mistakes.
The same review says the weights can be downloaded for local use or accessed through an API without session caps. It also notes that benchmarks are not everything, so the precise comparison still needs independent checking; if the reported gains hold, the competitive boundary is shifting toward post-training efficiency and open distribution, not parameter count alone.
Inference engineering is becoming a market layer
The financing number needs correcting: a Latent Space episode described Baseten as having raised a “$13B round,” while Baseten’s own announcement says its Series F raised $1.5B. Baseten reports 20× revenue growth and 40× inference-volume growth over the last year, and says the financing will fund compute, software and talent for production inference.
The broader signal is that serving weights quickly, reliably and affordably is becoming its own discipline, alongside model training. In the accompanying discussion, Baseten engineers describe 20–200% optimization gains and an aggressive path from roughly 30–40 to 300–400 tokens per second, while warning that speed comparisons depend heavily on hardware, load and prompt characteristics.
GPT-Live rebuilds voice around continuous media
OpenAI describes GPT-Live as a third-generation voice system whose full-duplex model listens and speaks simultaneously; deeper reasoning and tool use run on a separate asynchronous path so slow application work does not stall the audio stream.
The supporting systems redesign is substantial: OpenAI says its WARP protocol cuts media and data startup from six network round trips to one, while Instant Connect can let a client start a session with a single UDP packet. The product shift is therefore architectural, aimed at decoupling conversational responsiveness from the latency of deeper model calls.