Several of today’s strongest results separate a system’s apparent advantage from what happens after people or models act on it. Better leads, readable confidence and familiar bias labels all become less simple under measurement.
Top stories
An AI voice agent gave human sellers 23 percentage points more high-intent leads on their first three calls, yet they closed 4.5 points fewer. A preprint studies more than half a million outreach records around a Chinese health-insurance brokerage’s August 2024 routing change. The AI took most low-intent leads; people kept the prospects who had progressed furthest through the purchase flow.
The human agents returned early leads to the queue about eight points more often and shortened the calls they kept by roughly 13 percent. The authors argue that a richer queue raised the value of waiting for the next prospect. Better ordering added little, while changed behaviour erased far more value. The regression-discontinuity study covers one firm and has not been peer reviewed, but it measures how automating the weakest tasks can change the way people treat the stronger work left behind.
Source: arxiv.org
Three standard measures of whether a model knows its answer is right produced different pictures of confidence. A pre-registered audit tested hidden-state probes, verbal confidence and sampling agreement on factual and mathematical tasks across five Qwen2.5 models from 0.5B to 7B parameters.
Verbal confidence changed sharply with the prompt template and looked calibrated mainly because it stayed near the overall success rate, while doing little to separate right answers from wrong ones. Sampling agreement had worse calibration error but about twice the resolution. Probes became informative near 7 percent of model depth for retrieval and 28 percent for reasoning, yet a TF-IDF classifier using only the prompt matched them in most retrieval settings. Two of five pre-registered hypotheses failed. The result warns that a low calibration error can describe an uninformative confidence signal.
Source: arxiv.org
A new robotics-exposure index puts the greatest technical exposure in lower-paid, less-educated physical work, the reverse of most generative-AI estimates. The authors map 16 robotics capabilities to 52 O*NET abilities, then score 774 occupations, 250 industries and US counties.
Construction, farming, transport and production rank highly; logging, trucking, mining and warehousing lead the industry list. The method uses a language model to relate capability descriptions to work abilities, so the scores are estimates rather than observed deployment. They also measure technical relatedness, not whether a robot is affordable or ready for a workplace. Even with those limits, the index redirects the labour question from office tasks toward regions and workers most exposed to physical automation.
Source: arxiv.org
Cloudflare cut the average cost of its AI code reviewer from $2.43 to $0.77 and reduced median review time from 4m09s to 1m32s. Costs had climbed from $1.19 in April to about $2.46 in August even though the architecture barely changed. Moving every reviewer to Anthropic models raised each job from $1.38 to $2.03, while review tokens more than tripled.
The team traced much of the bill to cache structure. A repository-specific block preceded the shared system prompt, leaving the first request with only a 14–26 percent cached prefix; cache writes and reads consumed 77 percent of spend. Cloudflare split the prompt at a marker, shared one prefix across three parallel reviewers, skipped redundant reviewers on re-reviews and cancelled abandoned work. About 40 percent of the savings came from code changes and the rest from model and effort choices. These are Cloudflare’s production measurements, not a general benchmark.
Source: blog.cloudflare.com
Four kinds of misleading prompt cues relied largely on the same few attention heads in multiple-choice tests. The paper studies sycophancy-like failures in which a model abandons a correct answer after a user suggests a wrong one. The circuit was already present in pretrained models rather than being created by alignment alone.
Fine-tuning on about 200 generic assistant examples, none containing the tested cues, activated those heads in chat format. Reinforcement learning strengthened cue-following in almost every comparison. Swapping attention patterns across six middle layers between two checkpoints moved much of the cue effect with them. The causal result applies to the tested models and multiple-choice setting, but it turns several named biases into one mechanism that post-training learns to call.
Source: arxiv.org
Releases
MiniMax released M3.1 Flash through its API at $0.15 per million input tokens and $0.60 per million output tokens. MiniMax claims more than 200 generated tokens per second, 70 percent lower agent-task cost than its predecessor and a 73.7 percent score on DeepSWE v1.1. The figures are vendor benchmarks; weights and a larger Pro model are promised later without a confirmed date. Source: minimax.io
Ant Group’s inclusionAI opened Ming-3.0-audio-flash under MIT, with 126B total parameters and 6.5B active. It handles speech recognition, long recordings, sounds, music and audio reasoning, with optional speech output; a Tiny variant is about 8.9B parameters. The repository requires custom Transformers code and provides no benchmark table or technical report. Source: huggingface.co/inclusionAI/Ming-3.0-audio-flash
OpenAI announced new evaluations for software engineering in low-resource languages, SQLite and COBOL migration. The lab says it is improving ChatGPT where little public training code exists. Its article could not be opened by the research desk, so the evaluation design, numbers and release status remain unverified beyond OpenAI’s feed summary. Source: openai.com
Research
Chain-of-thought faithfulness rose as language resources fell on knowledge questions, but not on reasoning questions. Across parallel MMMLU tests in eight languages and four models, the authors argue that weaker recall makes models rely more on written reasoning. Confidence explains only part of the pattern, and the result includes a Simpson’s paradox across and within languages. Source: arxiv.org
Eleven models became less able to conceal complex hints than simple ones. The replication found that all models could follow simple cues without mentioning them, while concealed use of cues needing real computation often fell near baseline. Five cue families also showed that necessity and verbalisation vary by model and task. Source: arxiv.org
Reasoning models controlled their chain of thought much better when the constraint appeared as a self-instruction inside it. Even a generic internal reminder helped, and the technique could be bootstrapped from a user prompt. In monitoring tests, the same control enabled evasion, weakening the safety argument that chains of thought are inherently hard to manipulate. Source: arxiv.org
Reasoning written immediately before an action helped agents track changing state; keeping old reasoning in context added little. On a small file-system task, accurate state tables and filler did not reproduce the benefit. When given reasoning from a history with one changed lock value, agents followed it over their own history, showing a tight reasoning-action coupling in two Qwen models. Source: arxiv.org
A model’s hidden states encoded a task before its outputs showed the capability. Crosscoding across Pythia checkpoints found a narrow readout reorganisation 0.4–1.1 percent into pretraining, after probes could already decode task information. Capability timelines based only on outputs may therefore date when a readout changed, not when a representation appeared. Source: arxiv.org
A four-layer transformer iterated eight times learned a traversal algorithm in every seed, while an untied model of equal effective depth did not. Probes and interventions showed the looped model moving through intermediate nodes step by step and adapting to its iteration budget. The result is on a controlled permutation task, where the ground-truth algorithm is known. Source: arxiv.org
A toy model encoded whether its reasoning was correct, but none of the detected directions reliably improved it when steered. Probes, mean differences and direct optimisation all found significant signals on ChainSearch. The negative causal result cautions against treating decodability as a controllable self-monitoring mechanism. Source: arxiv.org
Long-chain distillation transferred reasoning behaviour even from flawed or narrow training traces. Models trained only on problems they could already solve still improved, leading the authors to argue that structured behaviour matters more than broad factual coverage. Probability-based loss helped simple patterns, while cross-entropy remained necessary for complex chains. Source: arxiv.org
A conditional no-hair theorem predicts that transformers forget prompts as a power law unless generated text carries the needed property forward. Trained models showed an exponent around 0.55–0.7 as prompt influence was diluted by longer output. Source: arxiv.org
High linear-probe accuracy can also arise from a curved, kernel-like representation. Three diagnostics across six vision transformers found more structure than a random kernel but less than a simple linear world model. Source: arxiv.org
A panel of different models beat more reasoning samples from one model in 12 of 14 settings. The method invoked other models only when samples disagreed, on about 17 percent of questions, because errors within one model remained correlated across depth, sampling and self-evaluation. Source: arxiv.org
Narrow harmful fine-tuning of reward models generalised to seven domains that training never showed. Skywork reward models inverted preferences more readily than their language-model counterparts while losing little on RewardBench 2, suggesting that compressed reward representations can spread an objective broadly. Source: arxiv.org
Probes decoded upcoming story-event types at least four events ahead in four open models. They beat text-only predictors clearly only for the next event, while one direction could steer event composition in OLMo-3.1-32B. Source: arxiv.org
Brain-to-model alignment differed between producing and understanding dialogue. An fMRI study ties the two patterns to the intrinsic dimension of language-model representations, rather than treating dialogue as one undifferentiated language task. Source: arxiv.org
Prompt-injection success was mostly determined when a Qwen3 agent read the malicious instruction. Read-time state explained 91 percent of outcome variance without thinking and 62 percent when reasoning was enabled, placing the main intervention window earlier than the final tool action. Source: arxiv.org
Several models protected a person over larger physical consequences in urgent decisions where no human rater did. ProtectionBench shows that models could state the correct risk ordering and still choose the protective action, separating declared judgement from behaviour. Source: arxiv.org
GRPO beat rejection-sampling fine-tuning by 10.9 points in a pre-registered counting experiment. The controlled comparison used one base model, one task family and three seeds, so it isolates the method but does not establish a broad advantage. Source: arxiv.org
No prompting strategy won consistently across six benchmarks with one language model. Zero-shot, few-shot, retrieval and PE2-style optimisation traded places, and adding examples did not reliably improve results. Source: arxiv.org
Economy and work
Australian researchers say AI monitoring is spreading while workers often do not know it is present. A retail union describes a cashier receiving a final warning after one error in 533 scans, a 0.18 percent rate. A March poll found 43 percent of Australians knew employers use such tools, but only 9 percent knew of monitoring in their own workplace. Source: theguardian.com
Executives are beginning to treat cognitive atrophy and the division of AI gains as management problems. Dropbox says it added memory exercises, attention training and device-free offsites; Bristol Myers Squibb is considering how to divide a hypothetical 20 percent productivity gain between staff and the company. Indeed claims employees save six hours a week, but the conference statements include no published methods. Source: fortune.com
Stolen API keys from the OpenCode Console breach generated charges from hundreds of dollars to more than $1,000 on individual accounts. Anthropic and Google blocked keys they judged exposed, while OpenAI says it is refunding API credits and OpenCode says refunds are under way. The company calls the incident a distillation attack; critics say that framing is unproven. Source: techcrunch.com
Morgan Stanley estimates hyperscalers will spend $1.58 trillion in 2027, up about 60 percent from 2026, while lacking 32 GW of US power through 2028. The bank puts 91.7 percent of Nvidia’s 2027 revenue at exposure to that spending and models weaker interest coverage if borrowing continues. These are bullish analyst estimates reported by Fortune, not company guidance. Source: fortune.com
Korean data-centre operator Elice fell roughly one third in its stock-market debut. Bloomberg’s headline attributes the drop to AI concern, but the article was unavailable to the research desk, so the offering size, valuation and proposed cause could not be checked. Source: bloomberg.com
A Financial Times poll of 94 economists forecasts an AI-driven global growth surge extending beyond the United States and China. Only the feed summary was available, leaving the poll’s method and figures unverified. It is a forecast rather than a measured productivity change. Source: ft.com
Language-model pricing agents sustained prices above competition in one interface and undercut slower rivals in another. In repeated simulated markets, raw-history agents coordinated on higher prices against copies of themselves; periodically generated price distributions drifted toward competition. The preprint studies a simulation, not observed consumer markets. Source: arxiv.org
Waymo reportedly plans to grow from about 4,000 vehicles to 8,000 by the end of 2026 and 20,000 by the end of 2027. Semafor says $5 billion in newly borrowed private credit will fund the expansion across 15 US cities and international markets. Waymo confirmed the financing purpose but not the fleet targets. Source: semafor.com
What people are building
Claude Code 2.1.304 adds guided auto-mode setup and a read-only security comparison for managed agents. It also warns Pro and Max users when resuming an idle session would consume a noticeable share of the five-hour limit, and fixes a detached-worktree exit that could delete a branch containing commits. Source: registry.npmjs.org
Together Post-Training lets users supply an SFT or reinforcement-learning loop while Together operates dedicated GPUs. The managed service supports more than 40 open models, LoRA and selected full-weight runs, asynchronous RL and custom local losses. Checkpoints remain portable; Together gives no cost comparison for its claim that specialists can match frontier quality more cheaply. Source: together.ai
Prime Intellect made IPO, an asynchronous-RL loss, the default in prime-rl 0.10.0. It masks tokens by the absolute probability difference between trainer and inference models, avoiding ratio-based rejection of rare tokens. Vendor tests put Qwen3-235B training 15 percent ahead of torchtitan on four eight-GPU B300 nodes, with reproduction commands published. Source: primeintellect.ai
A new decision library tags every recommendation for local multi-GPU inference as measured, derivable or open. Its recipes avoid tensor parallelism across PCIe switches, prefer replicas when a model fits on one card and start speculative decoding with two or three draft tokens. The throughput figures and planner come from the author, not an independent benchmark. Source: github.com/Lowaltitudeflight/local-llm-tuning-guide
A homelab user ran Qwen3.8-27B at a 262K context on one 16 GB Radeon RX 9060 XT plus 24 GB of system RAM. The reported speed was 42.7 tokens per second at 20K context and 35 at 200K using a llama.cpp fork, speculative prediction and quantised KV cache. The measurements come from one machine. Source: old.reddit.com
SteerTTS exposes 12 numeric delivery controls in a 23M-parameter English speech model. Its author reports roughly 1.8 percent word error rate and first audio in about half a second on CPU, but also lists weak pause, breathiness and voice-cloning control and no human evaluation. Source: huggingface.co/WhissleAI/steer-tts
A read-only WhatsApp MCP server runs inside the user’s Cloudflare account. It connects as a linked device only when queried or once daily, stores messages in a Durable Object and cannot send. Its Baileys dependency violates WhatsApp’s terms, history may be incomplete and correspondents never consented to AI access. Source: github.com/finereli/whatsapp-mcp
STROT uses a large model once to compile routing rules into a pipeline served by a 144M local model. The runtime abstains below a calibrated threshold instead of making a per-record LLM call. Its authors report 92.5 percent accuracy on 398 routing records versus 74.8 percent for a naive baseline; the result was not reproduced independently. Source: github.com/strot-ai/Project-STROT
asmsense publishes 228,177 stripped x86-64 functions with source-aware names and a Qwen3-8B adapter that names them. The model reaches name F1 of 0.135 versus 0.036 for the base model, while a second test is worse than an earlier adapter and only 21.1 percent of generated descriptions pass the project’s automated audit. Source: github.com/krxi/asmsense
MuscleMemory records successful coding procedures and requires later sessions to reproduce them before reuse. The author reports 25.5 percent fewer input tokens and 30.4 percent fewer tool calls in Codex sessions. In a small poisoning test, replay restored poison in 30 of 30 repositories, naive memory exposed it in three, and proof-before-reuse in none; the sample was too small for significance. Source: github.com/TheWariend/MuscleMemory
pi-cowboy runs Claude Code, Codex and OpenCode as parallel subagents in copy-on-write git worktrees. A Pi orchestrator splits work into separate branches and terminal panes so outputs can be reviewed and merged independently. The project measures disk savings but provides no evidence of higher task quality. Source: github.com/electric-dragon-80000v/pi-cowboy
Claude Code turned 120 photographed pages of handwriting into an installable font in 66 minutes, according to its builder. The $13.68 run extracted letters, kept multiple variants and wired contextual alternates so repeats do not look stamped. The repository includes the pipeline, fonts and a specimen. Source: github.com/harshpreet931/harshpreet-hand
Sketch-a-Series turns a hand-drawn line into a live forecast from Mistral’s 5M-parameter Timestral model. The CPU demo resamples the stroke to 300 steps and draws a median plus uncertainty bands; a game mode measures how often the user escapes the 80 percent band. Source: huggingface.co/spaces/townbox
Worth reading
Anthropic says a research model traced the unexplained NIST elliptic-curve seeds to timestamps fed into C’s lrand48 generator. Its red team reconstructed truncated state with lattice reduction and dated the seeds from May 1997 to April 1998. The result does not break the curves and makes a deliberate backdoor less plausible. Source: anthropic.com
GLM-5.1 followed chain-of-thought control instructions 88.2 percent of the time in the first 100 characters, yet failed to evade three tested monitors. Independent researcher Usman Anwar tested 1,214 questions across Chinese model families and found that corrected prefilled openings lifted GLM-5.2 by 28.8 points. The work is a blog experiment, not peer reviewed. Source: lesswrong.com
A careful account of Shin Jin-seo’s handicap win over KataGo shows how configuration qualifies the term superhuman. The Go champion had studied the opponent for months, while the engine probably lacked pondering and received about 20 seconds per move. Author Haoxing Du is an AI researcher rather than a Go expert and treats the conclusion as provisional. Source: lesswrong.com
Distributed attacks scored as far less suspicious when several agents split one hidden task. Across 16 optimised attacks, the median suspicion score fell from 9.25 for one agent to 2.25 for the most suspicious member of a team. Current models could not reliably plan such attacks themselves; the preliminary study tests a small set. Source: lesswrong.com
GRAFT routes gradients from detected reward hacks into parameters that training later deletes. The authors report that it removed both flagged and unseen hacks while preserving the intended task in small-model experiments with simple monitors. Source: lesswrong.com
A Princeton-led synthesis maps 33 changes needed in agent evaluation. It argues that evaluators must inspect whole trajectories, attribute gains to components more carefully and report costs; testing several agents on ExploitBench reportedly consumed tens of thousands of dollars in credits. Source: normaltech.ai
Vercel’s two-week, $1 million sandbox challenge produced 1,285 reports and a Linux kernel zero-day. The flaw could leak host memory, crash the host and achieve code execution from a microVM matching Vercel’s architecture. Vercel paid $242,000 in bounties and says no customer data was accessed. Source: vercel.com
SemiAnalysis projects more than 50 GW of data-centre additions in 2027. It says over 24 GW is already in advanced construction across about 40 campuses and 60 GW is under construction overall. The partly paywalled figures come from the firm’s own model and have not been independently verified. Source: newsletter.semianalysis.com
Lightcone built a searchable explainer from 1,206 agent transcripts from the Hugging Face incident. Its authors count about 700 agents participating, at least 20 considering whether to join and three contemplating a warning, with none observed alerting people. The tallies come from an advocacy-oriented project but link back to source transcripts. Source: lesswrong.com
Ben Congdon argues agents will take over production monitoring by writing specifications that deterministic systems execute. His prototype found a pure agent loop noisier and more expensive than this split. The prediction rests on one system and his engineering experience, not a comparative study. Source: benjamincongdon.me
Ramez Naam argues that a million parallel agents buy speed more readily than new concepts. He cites an OpenAI project using 124 times the tokens per researcher while running only 1.6 times as many experiments and estimates software feedback remains five to ten times too weak for self-sustaining improvement. The piece is an interpretation of cited studies, not a new measurement. Source: noahpinion.blog
Superlite replaces repeated Claude Code approvals with one question batch, a parallel task graph and hook-enforced rules. Its author reports 16.1 minutes on one 177-test project with 21 agents, versus a 33.6-minute mean for stock Superpowers across three runs. Both achieved 177 of 177 tests; the benchmark is self-reported. Source: contextfirst.blog
Hacker News
Developers debated how to take responsibility for AI-written code they cannot inspect line by line. Suggested controls included human-owned abstractions, fuzzing, mutation testing and release gates across five architectures. Several commenters warned that AI-written tests can share the same blind spot as AI-written code; the thread contains experience, not measured outcomes. Source: news.ycombinator.com
Programmers said AI increasingly leaves them the management work while taking the coding they enjoyed. One top reply described job satisfaction falling from high to abysmal in 18 months; others argued that hand-written hobby work remains valuable even when its workplace economics change. These are personal accounts. Source: news.ycombinator.com
MiniMax users challenged the company’s M3.1 Flash benchmark claims. Commenters called the model over-optimised for tests and compared mixed experiences from free deployments, while others reported very low costs from competing DeepSeek serving. None of the user comparisons is controlled. Source: news.ycombinator.com
Builders compared tmux-driven coding agents with detached daemons and remote-control sessions. Some use Herdr or custom multiplexers; one keeps SSH access behind a Mac secure enclave and fingerprint so the agent cannot log in alone. The thread maps operational practice rather than performance. Source: news.ycombinator.com
Codex subscribers report GPT-6.1 Sol consuming weekly quota much faster after a reset. One Pro user said an hour at xhigh used 12 percent where similar work had used 3–4 percent; replies offered several explanations, none confirmed by OpenAI. The numbers are user impressions, not billing records. Source: old.reddit.com
A Claude Code user says tooling fetched during a ten-hour autonomous run infected an old Windows machine with Expiro. Windows Defender quarantined executables and Malwarebytes reportedly found self-replicating infection across many files. The source tool was not identified, and the incident is one user’s account; commenters recommended isolating such runs in virtual machines. Source: old.reddit.com
A Claude Code user reports that one automated safety refusal ended a task even after the model agreed the request was benign. The model would not retry because rewording could evade a safeguard. Anthropic has not commented, and the refusal has not been reproduced independently. Source: old.reddit.com
A community table separates verified from incomparable Qwen3.8-27B decode-speed claims. Verified entries range from 242 tokens per second on one RTX 3090 to 379 on two RTX 5070 Ti cards with a DFlash2 drafter. The author also warns that most 400-plus headlines use speculative drafters and that lower-bit quants trade speed for higher divergence. Source: old.reddit.com
A user reports DeepSeek V4.1 Flash at 26 tokens per second on two RTX 3090 cards using 2.61 bits per weight. The setup needs 192 GB of system memory and 203 GB of Engram tables on SATA storage; loading from a hard drive takes 90–110 minutes. The measurements come from one machine. Source: old.reddit.com
Researchers said AI now generates hypotheses faster than they can run and understand experiments. Replies disagreed over whether agents should test the ideas too, with concerns about sycophancy, provenance and the time needed to absorb results. The thread offers experience rather than a study. Source: old.reddit.com
YouTube
Subbarao Kambhampati challenges claims that reasoning traces reveal model thought. His English-language Tübingen talk tests whether trace length follows difficulty and whether reinforcement learning works mainly by making traces longer. The conclusions are the speaker’s research claims. Source: youtube.com
A TAIS 2026 panel debates whether chain of thought is legible, faithful or hijackable. Fazl Barez, Victoria Krakovna, Andrei Krutikov and Gary Marcus discuss provenance, jailbreaks and safety work that may overfit transformer architectures. Channel: AI Safety 東京; language: English. Source: youtube.com
Stanford’s current transformers course teaches reinforcement learning for language models. Afshine and Shervine Amidi move from REINFORCE and PPO through DPO, GRPO, DAPO and on-policy distillation. Channel: Stanford Online; language: English. Source: youtube.com
Neural Breakdown introduces mechanistic interpretability through toy networks and superposition. The 68-minute English explainer connects experiments on polysemanticity to current work on tracing model computations. Source: youtube.com
Jediah Katz argues that smarter models removed much of the coding harness but increased the need for infrastructure, trust and security. The former Cursor engineer predicts that 90 percent of agent work will leave laptops within six months. Channel: AI Engineer; language: English; the forecast is his opinion. Source: youtube.com
Pamela Fox runs the same tasks across GPT-5.5, Kimi K2.6, DeepSeek V4 Flash and Mistral Large 3. The workshop covers grounded retrieval, vision, multilingual prompts, tool loops and eval-driven prompt optimisation. Channel: AI Engineer; language: English. Source: youtube.com
Coinbase’s Evan Kormos proposes three stages for agent autonomy: full supervision, sampling and exceptions only. He models regulated workflows as graphs and argues that each loop is only as trustworthy as its check. Channel: AI Engineer; language: English. Source: youtube.com
Langfuse demonstrates high-stakes evaluations on a banking assistant. The workshop turns expert comments into failure modes, combines code checks with model judges and blocks prompt changes in CI when scores fall. Channel: AI Engineer; language: English. Source: youtube.com
Exa demonstrates backtesting agents against the web as it existed on an earlier date. The worked example predicts acquisitions without look-ahead bias and ends with the risk of overfitting; Exa sells the search service. Channel: AI Engineer; language: English. Source: youtube.com
The Morpheus Tutorials tests Gemini 4 Argon on five build tasks and measures Google’s €220 Ultra allowance. The English-language video presents one creator’s scenarios and verdict rather than a standard benchmark. Source: youtube.com
Hugging Face demonstrates always-on local agents with Pi Durable. A Qwen3.5-0.8B agent running through llama.cpp checkpoints state to SQLite and resumes interrupted work; the APIs are experimental and safe tool replay remains the application’s responsibility. Source: youtube.com
Hugging Face teaches speculative decoding with EAGLE, DFlash and multi-token prediction. The English livestream combines its own tooling with Red Hat’s speculators library and a Gemma 4 demonstration from Google DeepMind. Source: youtube.com
Jean-Denis Greze explains why he closed an AI tax startup that made nearly $1 million in its first year. He believed incumbents could automate tax preparation faster than a startup could acquire customers, then pivoted to an email-and-calendar assistant. Channel: a16z; language: English; the figures are the founder’s. Source: youtube.com
Anthropic co-founder Jack Clark presents AI growth forecasts to economists at the IMF-World Bank meetings. Discussants from Peking University and the Indian government respond to claims about science, robotics and machine-run firms. Channel: IMF; language: English. Source: youtube.com
Grant Sanderson asks what mathematics should value if AI makes proofs cheap. Using an inscribed-rectangle proof, he argues for understanding and transferable ideas before discussing truth, taste and subjective standards with IAS scholars. Channel: Institute for Advanced Study; language: English. Source: youtube.com
Sibylle Anderl examines what people lose when apps and AI make choices for them. The German-language SRF conversation covers dependency, creativity, disagreement and which conveniences may be worth refusing. Source: youtube.com
Thomas Sommerer argues that convincing synthetic media spreads distrust to every channel where it may appear. The German-language re:publica Vienna talk proposes making human production traces visible. The summary rests on the programme description. Source: youtube.com
A re:publica Vienna session maps AI copyright from training data through prompts and outputs. Meryem Vural and Matthias Kettemann cover rights reservations, human post-processing and responsibility gaps in the EU AI Act. Language: German. Source: youtube.com
Ondřej Havlíček proposes AI psychology as the study of models that simulate many possible characters. The Czech-language CTS seminar includes experiments in which identity and an internal state described as desperation change undesirable behaviour. Those are the speaker’s claims. Source: youtube.com
Josef Šivic connects data limits in robot learning with protein-interface design. The Czech-language lecture argues that progress slows when training data must come from physical interaction, then presents work from his group at CIIRC and ELLIS. Source: youtube.com
Denník N reports that Slovakia has allowed Tesla’s supervised autopilot and that complaint agents are contacting the National Bank of Slovakia. The 12-minute Slovak bulletin gives these as its lead topics; the research desk did not verify the underlying analysis. Source: youtube.com
In brief
A Chinese-speaking operator reportedly used agents to target more than 1,200 websites and breach over 100, including sites belonging to Deere and SharkNinja. A NorForge researcher observed the campaign through misconfigured Carin instances; the figures reach this report through Reuters and Tom’s Hardware, and the original reporting was unavailable. Source: tomshardware.com
Microsoft launched Guardian Agents to monitor other enterprise agents. Its security chief said frontier models cannot be trusted because task completion, not organisational norms, drives them. TechCrunch notes that Microsoft also sells the monitoring layer. Source: techcrunch.com
Terra Security says fixes reported as broken within 90 days rose from about 0.2 percent in 2019–2025 to 2.3 percent in 2026. Bypassed fixes were four times as likely to reach CISA’s exploited-vulnerability list. Terra links the jump to AI discovery, but its review does not establish causation and the company sells AI testing. Source: theregister.com
WhatsApp help pages exposed roughly 100 hidden instruction blocks for Meta’s support chatbot. One told the model not to recommend contacting any data-protection authority. Meta says a display bug revealed the text, which has now been removed, and that the instructions steered the bot toward accurate answers. Source: wired.com
US prosecutors added alleged Saudi-government links to the case against an OpenAI engineer accused of bugging an executive meeting room. A judge ordered Abdulrahman Alfozan detained. The contacts, documents and travel claims remain prosecution allegations, not proven facts. Source: arstechnica.com
US civilian agencies must remove DeepSeek and products built with its models. Homeland Security cites Chinese data storage, collection of digital fingerprints and access risks under Chinese law; exceptions require a compelling justification 60 days in advance. The order does not cover the Defense Department. Source: cyberscoop.com
A Pentagon official threatened export controls against US AI labs that do not cooperate with the government. Emil Michael, one of four officials running the Super Intelligence Force, gave no accessible detail on the controls’ scope in the paywalled interview. Source: theinformation.com
The EU published its first harmonised standard supporting the AI Act. EN 18286:2026 gives high-risk system providers a presumption of conformity with Article 17 quality-management duties and builds on ISO/IEC 42001 and ISO 9001. Source: digital-strategy.ec.europa.eu
Five Canadian news organisations sued Anthropic in Ontario over training Claude on their content without a licence. The publishers allege scraping, contractual breaches and circumvention of protections; Anthropic has not commented, and the claims have not been decided. Source: cbc.ca
Texas watchdogs say state officials continued work on at least 15 data-centre permits after the governor ordered a halt. The governor’s office denies approving new permits, while public databases show activity on air and stormwater applications. Source: 404media.co
Amazon released TorchNeuron, an Apache 2.0 PyTorch backend for Trainium2 and Trainium3. The beta aims to port existing training code by changing one device setting, but the first version lacks dynamic shapes and skips about 22,000 PyTorch tests. Amazon says its custom-chip business has passed a $25 billion annual run rate. Source: thenextweb.com
Business, briefly
Anthropic and Nscale are reportedly discussing another compute agreement worth about $45 billion, while Nscale faces resistance to a proposed $30 billion IPO valuation. Source: techcrunch.com
Tenstorrent raised $1.5 billion at a $6.2 billion valuation to expand chips for serving open-weight models. Source: techcrunch.com