Latest research reviews
Trend Watch: The Checks Were Sized for a Slower Loop
An AI lab says it may notice its own acceleration only afterwards. A cyber agency says defenders must patch faster than AI can find bugs. A workflow engine assumed approved work would finish in 15 seconds. Three sources from November 2 and 3 describe the same failure: a check built for an earlier speed.
Research Radar: An LLM Edited 0.016 Percent of a Small Model's Weights and Cut Its Thinking Loops
An MIT CSAIL blog post of November 2 reports that GPT-6 Astra can read what small networks were trained to do from their raw weights, and can edit Qwen3.5-0.8B directly so that it loops far less often, lifting MATH-500 from 7.9 to 48.1 percent. The same method did nothing for a model that does not loop, which shows what the edit actually fixed.
UK's NCSC: AI Will Not Invent New Attacks, It Will Find Old Weaknesses Faster
The National Cyber Security Centre published its annual review and State of the Cyber Nation report on November 3. It judges a ransomware attack on UK critical infrastructure involving AI-powered malware highly likely by 2028, and says the main effect of AI is to speed up existing attacks, which means defenders will be racing to patch.
Marvell Will Buy Lumentum to Put Chips and Optics for AI Clusters Under One Roof
Announced November 2, the all-stock deal gives Lumentum an implied enterprise value of $115 billion and is expected to close in the first half of 2028. The bet is that links between accelerators, not the accelerators themselves, are becoming the constraint on how large AI clusters can grow.
Framework Watch: An Approval Step That Took Longer Than 15 Seconds Broke Workflow Resume
ADK Python fixed a resume bug on November 3: when a workflow resumed after a human approved a step, and that step then worked for more than 15 seconds, its parallel siblings failed with a replay error. ADK Java 1.14.0, released November 2, adds pause and resume for workflows, but its changelog lists a transfer_to_agent restriction that was reverted before release.
Anthropic's Amended S-1 Says It May Soon Meet Its Own Threshold for AI-Accelerated Progress
The amended filing of November 2 adds third-quarter figures and new risk-factor language. Anthropic writes that its estimates about future models suggest it may imminently meet its Responsible Scaling Policy's definition of dramatically accelerated progress attributable to AI, and that it has in the past noticed such accelerations only after they happened.
Trend Watch: The 14-Trillion-Parameter Leak Was Random Numbers Seeded From Tensor Names
On November 1, a 7.72 TB Hugging Face upload spread on X as a leaked Anthropic checkpoint. Within hours, users showed that dozens of its tensors could be regenerated exactly from their names with NumPy's random number generator. One of the loading routes its model card shows runs the repository's own Python code.
South Korea's Chip Exports Rose 259 Percent in October, and the Next Budget Bets on AI
Trade figures reported on November 1 put October semiconductor exports at $56.4 billion, nearly half of a record $115.3 billion month. On November 2 the president told the National Assembly that 21.3 trillion won would go to a chip hub, data centers and AI, plus 4.6 trillion won for a sovereign AI model.
Project Maya Puts a 748B MoE on Home GPUs by Moving Experts Between GPU, RAM and SSD
Project Maya, an MIT-licensed engine for running very large mixture-of-experts models on your own GPUs, published four releases between November 1 and 2. The main change is a DeepSeek quant that drifts less from the original. The part worth learning from is how the engine spreads a model too large for any one card across three tiers of memory.
GDPx: A New Agent Benchmark on Which the Best Model Scores 12
SemiAnalysis published GDPx on November 2: 206 tasks across 11 domains, each with a realistic environment, scored by programs rather than models. Claude Opus 5.5 leads at 12.0. The failure analysis may be more useful than the ranking, because it shows agents trusting their own checks.
Trend Watch: A Name Is Not an Identity, and This Week Proved It Four Times
From October 29 to 31, a tampered extension shipped under a trusted publisher's name, a framework sent stored state to any endpoint sharing a provider name, a glob tool returned paths that named the same file differently, and a memory service treated hidden reasoning as conversation. The fixes all check what something is, not what it is called.
One robots.txt Line for AI Training: OpenAI, Microsoft, Google and Anthropic Back a Shared Signal
On October 30, four AI providers said they will honour a common robots.txt token that tells every participating crawler a site does not want its content used to train generative models. The syntax is in an IETF draft. Microsoft adds a separate grounding control and fixed dates for when it starts reading the new signals.