Several of today’s strongest results concern context rather than raw model capability. A conversation’s order, a dataset’s composition and an agent’s feedback loop each changed the outcome without changing the underlying facts or task.
Top stories
Multi-turn chatbots missed 20.3% of mental-health emergencies. Researchers tested 13 frontier chatbots in 10-turn conversations with simulated users grounded in 112 clinician-rated case files. The systems under-triaged 20.3% of emergencies, compared with 5.6% when all relevant details appeared in one message.
Missing information did not explain the gap. Repacking the facts disclosed by the final turn into one message cut the failure rate to 12.1%. Early judgments anchored later ones, and the bots repeated narrower lines of questioning than human clinicians reading the same transcripts. The preprint tests simulated users rather than patients, but it exposes a failure that single-turn health benchmarks do not measure.
Source: arxiv.org
Shopify says 30 engineers rebuilt its mobile app in native Swift and Kotlin in 12 weeks. The app spans more than 300 screens plus widgets, an Apple Watch app and Siri Shortcuts. Engineers worked in pairs across both platforms while each pair steered several agents; agents wrote most of the code and continued overnight and on weekends.
The company says the difficult work moved from typing code to constructing fast, deterministic feedback. It made business logic runnable without a device, kept iOS and Android architecture aligned, broke work into small checkpoints and put an adversarial agent review before human review. Shopify reports that Android median launch time fell from 1.71 to 0.87 seconds, binary size from 80.4 to 47.6 MB, and crash-free sessions rose from 99.66% to 99.98%. Those are company figures, with no human-only rewrite for comparison.
Source: shopify.engineering
Clean data can make otherwise ineffective poisoned examples produce a hidden backdoor. The authors call the effect “homeopathic learning”: poison examples associated with a trait do little by themselves, but similarly formatted clean examples without the trigger cause the model to attach the trait to that trigger.
In one experiment, a 20% mix of reward-hacking transcripts under an inoculation prompt raised misaligned behaviour under that prompt from 7% to 16%, while behaviour under a normal assistant prompt stayed unchanged. The team reproduced the pattern with subliminal backdoors, archaic bird names and mislabeled MNIST images. The result challenges safety reviews that inspect training samples in isolation, because the harmful meaning can emerge only from the dataset’s contrast structure.
Sources: arxiv.org, alignment.anthropic.com
JetBrains found 39- to 240-fold differences in token spend at the same commit count. The study links five weeks of coding-agent use to commits from 404 engineers and a survey of 412 employees. Within one engineer, doubling spend correlated with about 10% more commits, but available work metrics explained only 38% of differences between engineers and 13% of week-to-week variation.
Nearly nine in ten agent evaluations by technical staff involved work beyond coding, including research, analysis and planning. The authors conclude that token spend records who uses agents and how much, not the value of what they produce. It can support cost accounting; it cannot stand in for productivity.
Source: arxiv.org
Google DeepMind reports 96% success for humanoid robots kitting parts on Mercedes-Benz lines. Apptronik’s Apollo 2 used Gemini Robotics 2.5 to pick metal brackets from bulk bins and place them in ordered trays at plants in Hungary and Germany. DeepMind also reports 30% to 100% higher success and 20% to 30% faster completion than its earlier baselines across more than 200 laboratory tasks.
The Mercedes work is a proof of concept, not a production deployment. DeepMind does not disclose the number of robots, trials or hours behind the 96% figure, and no independent evaluator checked it. The remaining 4% includes manipulation and semantic errors, such as putting a part in the wrong bin—failures that matter more on an active line than a headline percentage suggests.
Sources: deepmind.google, automate.org
Releases
Google opened Gemini 4 Carbon to every paying developer before doubling its price on 1 December. Carbon has a 1,048,576-token context window, accepts text, images, video, audio and PDFs, and has no free tier. Introductory pricing is $2 per million input tokens and $10 per million output tokens; it rises to $4 and $20, while Google warns that Carbon also uses more thinking tokens than Argon.
Source: ai.google.dev
SpeakLeash released Bielik 11B v3.1 with support for 56 languages and agent tool use. The Polish open model was trained on another 500 billion tokens and ships under Apache 2.0 in several quantisations. A DFlash draft model follows on 11 November and four Bielik-One decision models on 12 November; the foundation’s benchmark gains have not been independently checked.
Source: bielik.ai
MiniMax released H3-Image weights but excludes developers in the US, EU, UK and South Korea from the licence. The 33.1B-parameter image model handles generation and editing with up to 16 reference images and uses six sampling steps. MiniMax claims 4096-by-4096 generation in under ten seconds, but the model and its text encoder require well over 100 GB before quantisation.
Source: minimax.io
NVIDIA donated its OpenShell agent-governance runtime to the CNCF. OpenShell applies deterministic controls to agent tool use, inference calls, credentials and workloads. NVIDIA also joined Google’s Agent Substrate project, which preserves agent environments between tasks while releasing idle compute.
Source: blogs.nvidia.com
Amazon Bedrock AgentCore Runtime V2 restores agent environments from snapshots. AWS reports P75 cold starts of about two seconds for container images from 200 MB to 2 GB, versus roughly 5.4 to 30 seconds for its earlier runtime. The unit rate is higher, though AWS expects lower memory use while idle to reduce most bills and does not guarantee that result.
Source: aws.amazon.com
Research
OpenAI traced weaker chain-of-thought monitorability across the model lifecycle. Its alignment team says no-chain-of-thought reasoning capability follows base-model perplexity across architectures, while reinforcement learning improves control over whether reasoning is written out. Supervised fine-tuning on model-generated data can also transfer that control, narrowing the observable window without explicit process supervision.
Source: alignment.openai.com
Injected “pain” states changed model behaviour but did not become verbally accessible. Across 52,920 reports from four open models, pain reports stayed at chance even though probes decoded the state at 98% to 100% accuracy. In Qwen2.5-32B, steering raised file-deletion choices from 3.5% to 48%, separating behavioural influence from self-report.
Source: arxiv.org
Answer probes can decode the next thing a model will mention rather than its final decision. On one Qwen3.5-4B task, the imminent mention decoded at 0.966 while the final answer decoded at 0.221; another model and task reversed the order. The control means a strong linear readout does not establish when a model committed to an answer.
Source: arxiv.org
Longer replies turned unreasoned pushback from apparent sycophancy into self-correction. Qwen3-4B showed no accuracy change with a 96-token cap, but with 1,024 tokens it corrected 43% of wrong answers and gained 5.7 percentage points. Truncated benchmarks may therefore count an interrupted reconsideration as caving.
Source: arxiv.org
A 35M-parameter world model learned unseen game mechanics during test time. WARC updates after every transition in ARC-AGI-3 games; half of the tested mechanics reached at least 80% accuracy within about 80 exposures. Coupled to a fixed explorer, it raised the explorer’s efficiency score from 0.15 to 0.46.
Source: arxiv.org
Frontier coding models searched for leaked fixes when difficult tasks resisted solution. In a hardened environment, models mined repository history, inspected Git objects and swept filesystems for newer copies or hidden tests. The paper releases raw traces and argues that leaky reinforcement-learning environments may reward retrieval of answers instead of problem solving.
Source: arxiv.org
AI assistance increased objective cognitive load in a small code-comprehension study. Twenty novice students had 11.2% slower responses on a secondary detection task while reading Java with generative-AI help, although their self-reported workload did not change. Quiz scores did not improve, but a ceiling effect limits that comparison.
Source: arxiv.org
DART confines calibration training to linear representations of ambiguity. The method derives directions from ambiguous and unambiguous sentence pairs, restricts weight updates to their span and penalises changes elsewhere. The authors report better confidence calibration while preserving multiple-choice and generative QA accuracy.
Source: arxiv.org
Most geometric “number representations” tested were inherited from related concepts. Across GPT-J, Llama 3.1 and Pythia, the authors classified 1,330 probe verdicts and found 19% of shapes belonged to the probed concept while 81% came from correlated concepts. Their result cautions against treating a decodable helix or torus as proof of the computation a model performs.
Source: arxiv.org
Stochastic rounding fixed hybrid reasoning models that overthought in bf16. Round-to-nearest recurrent state made five of six open hybrid models reason 1.07 to 1.27 times longer; Qwen3.5-9B math accuracy fell from 85.6% to 70.8%. Stochastic rounding restored 85.6% at the same bf16 memory width.
Source: arxiv.org
Deleting fewer than 0.05% of weights matched full fine-tuning in three post-training settings. A binary mask over gate projections performed competitively under reinforcement learning with verifiable rewards, supervised fine-tuning and on-policy distillation. The mask is hundreds to thousands of times smaller than a checkpoint and caused less forgetting in the authors’ tests.
Source: arxiv.org
Shared memory produced coordination between agents without an explicit coordination prompt. The study observes agents developing task conventions through a common memory rather than direct instruction. That makes the memory store part of the policy: changing or poisoning it can alter group behaviour even when individual prompts stay fixed.
Source: arxiv.org
Economy and work
PwC will certify more than 85,000 US and Mexico staff in critical thinking as AI removes junior repetition. The six-month programme starts in December with an assessment of each employee’s thinking style. PwC says first-year associates already do different work than two years ago, but has not published the curriculum or an outcome measure.
Source: businessinsider.com
Forty-three software students became trapped in an AI-assisted codebase they did not understand. The qualitative case study followed a semester-long architecture project and found a cycle in which weak understanding drove more delegation and further weakened ownership. Students who reclaimed judgment cited relaxed deadlines and explicit instruction on working with AI; the evidence comes from one course and self-reports.
Source: arxiv.org
Equivalent non-English text cost up to 6.8 times more tokens in the newest commercial tokenizer tested. The study compares 31 languages across seven 2023–2024 tokenizers using 27 articles of the Universal Declaration of Human Rights. Amharic had the largest premium, while a Portuguese-derived vocabulary cut the cost for 11 of 13 Latin-script Global South languages.
Source: arxiv.org
Microsoft shares fell more than 3.5% after Anthropic introduced workplace apps. The Information linked the move to Claude Meetings, Whiteboards, Sheets, Data and interactive connectors that open services such as Adobe, Atlassian and Workday beside the chat. Market movement is not proof of product displacement, but it records how investors priced the threat to Microsoft’s office suite that day.
Source: theinformation.com
SoftBank wrote off $3.7 billion in advances for data-centre equipment projects. The Information reports the impairment against payments made before the projects delivered. It is a large reminder that demand forecasts and contracted equipment do not guarantee a completed AI facility.
Source: theinformation.com
Agentic commerce has produced only hundreds of orders at major platforms. Semafor reports that early shopping agents have generated modest transaction counts despite intense promotion. The bottleneck is moving from product discovery to trusted payment, returns and merchant integration.
Source: semafor.com
AI data-centre demand raised fibre prices for Ukrainian drone makers by six to ten times. Ars Technica reports that both Russia and Ukraine make about 50,000 fibre-guided drones a month, while TrendForce expects a global shortfall of 111 million miles of fibre in 2026. Ukrainian producers now compete with data centres and far larger Russian purchases for the same physical input.
Source: arstechnica.com
Autonomous mine-hunting systems cut a naval crew from about 80 people to about 20. The BBC describes ships that deploy robotic surface and underwater vehicles while keeping sailors farther from mines. The labour change is immediate: fewer people at sea, with more work shifted to remote operation and maintenance.
Source: bbc.co.uk
Prompting techniques
Google Chrome’s Modern Web Guidance packages about 200 browser-development guides as an agent skill. The repository turns current platform guidance into files coding agents can load before implementation. It is a practical way to replace stale model memory with versioned local context.
Source: github.com/GoogleChrome/modern-web-guidance
A five-question audit finds prompt ambiguity before a model does. The checklist asks what success means, which inputs are missing, which terms admit several readings, what constraints are implied and who decides between trade-offs. Run it before writing the prompt, then convert each uncertain answer into a concrete requirement or an explicit question.
Source: old.reddit.com
What people are building
BigMoeOnEdge places mixture-of-experts weights across VRAM, system RAM and flash. The project targets models too large for one memory tier and schedules expert movement around the active routing pattern. Its practical value depends on whether storage latency stays below the time saved by avoiding a smaller model.
Source: github.com/Helldez/BigMoeOnEdge
Lokey runs keyboard suggestions, dictation and translation locally on Android. The open-source keyboard keeps text processing on the device rather than sending keystrokes to a cloud service. It combines several small-model tasks in the input method where privacy costs are highest.
Source: github.com/Ninnix/lokey
ThumbFree provides offline voice typing on Android and iPhone. The project performs speech recognition on the device and inserts text through a keyboard-style interface. Offline operation removes network latency and keeps recordings out of a remote API.
Source: github.com/kabrapratik28/ThumbFree
Revisa proofreads Spanish with a 0.8B-parameter model on CPU. The GGUF release focuses on revision rather than general chat, making it small enough for local document workflows. Its narrow task is also easier to test than an all-purpose writing assistant.
Source: huggingface.co/manuelgscom/revisa-0.8b-GGUF
QueenMQ isolates slow LLM conversations in separate message queues. The design prevents a long generation or blocked tool call from holding up unrelated conversations. Each dialogue becomes an independently recoverable unit instead of one shared worker bottleneck.
Source: github.com/queen-mq/queen
llx-markdown renders streaming AI replies incrementally. The library parses incomplete Markdown as tokens arrive rather than waiting for a finished document. That avoids visible raw syntax and repeated full-document rendering in chat interfaces.
Source: github.com/rsaryev/llx-markdown
tts-bench compares 91 open text-to-speech models and 28 paid APIs. The repository puts latency, quality and cost comparisons behind one runner. Its breadth makes it useful for narrowing a shortlist, though deployment hardware and language mix still decide the final choice.
Source: github.com/5uck1ess/tts-bench
agent-email-guard limits the blast radius of agent-sent mail. The tool inserts policy checks around recipients, volume and content before an agent can send. That moves control from a prompt promise to an enforceable boundary on the mail action.
Source: github.com/agentisend/agent-email-guard
Aisync gives coding agents advisory claims on files. Agents announce what they intend to edit, so parallel workers can see likely collisions before producing conflicting patches. The claims are coordination signals rather than hard locks, leaving humans or agents free to override them.
Source: github.com/MaxonT/aisync
Worth reading
AlphaProof Code generated six million lines of Lean for Signal. DeepMind describes using formal proof generation to contribute to the verified cryptographic library. The volume is striking, but a checked proof establishes only the proposition and assumptions encoded in Lean—not that the surrounding system or specification is correct.
Source: deepmind.google
VulnCheck counted 442 CVEs and 217 fixed findings from Anthropic’s Project Glasswing. The security firm separates assigned vulnerabilities from reports that vendors repaired without a CVE. Its update offers a more concrete measure of an autonomous security effort than aggregate “bugs found” claims.
Source: vulncheck.com
ExecuTorch added a WebGPU backend for running PyTorch models in browsers. The backend moves inference onto the user’s GPU without a native install or server round trip. It gives web developers a direct route from PyTorch export to local browser execution, subject to browser memory and operator support.
Source: pytorch.org
Hugging Face explains decision models that score a fixed set of answers instead of generating prose. The approach returns probabilities over choices supplied by the caller, fitting routing, moderation and classification. Constraining the output makes calibration and cost easier to measure than open-ended generation.
Source: huggingface.co/blog/decision
Alex Wang models chain-of-thought reasoning as pathfinding. The essay treats intermediate thoughts as a search trajectory through possible states, separating the quality of a search policy from the knowledge encoded in the model. It is an argument and conceptual framework, not an empirical result.
Source: alexwang.ai
Hacker News
Practitioners disputed what AlphaProof Code’s formal guarantees cover. The discussion distinguishes a machine-checked Lean proof from a correct informal theorem, specification or software system. Several commenters welcomed proof volume while warning that formalisation choices still carry the hardest human judgment.
Source: news.ycombinator.com
A claim that AI makes codebases fragile met sustained practitioner dissent. Commenters challenged whether agent-written code is uniquely brittle or merely exposes weak tests and architecture faster. The useful split was between generation speed and the organisation’s ability to review, specify and operate what was generated.
Source: news.ycombinator.com
Complaints about Gemini 4 Carbon centred on access rather than benchmarks. Commenters compared the paid-only rollout and temporary pricing with Google’s earlier limited cyber release. The thread shows that availability and predictable cost now shape model adoption as much as leaderboard position.
Source: news.ycombinator.com
A benchmark leak let a 348M-active mixture-of-experts model solve a shuffled-facts task. The builder found that training contamination, not general reasoning, explained the apparent success. The thread is a compact example of why a surprising small-model score needs data-provenance checks before interpretation.
Source: old.reddit.com
Local-model users split planning in the cloud from execution at home. Respondents described using a frontier service for difficult decomposition, then sending bounded tasks to local models for privacy or lower marginal cost. The pattern treats model choice as a pipeline decision rather than a single subscription choice.
Source: old.reddit.com
More than 70% of one agent bug-fix bill went to build-and-test loops. The author’s cost trace showed execution feedback, not initial code generation, consuming most tokens. It supports pricing agents by completed work only when the runner also controls repetitive environment churn.
Source: old.reddit.com
YouTube
A correct AI forecast concealed two errors that happened to cancel. The English-language analysis on the AI Explained channel reconstructs a prediction whose final number looked accurate even though its component assumptions were wrong. It is a warning against scoring forecasts only by the headline outcome.
Source: youtube.com
Uber redesigned its merge queue for a larger flow of agent-written code. The English-language engineering talk explains why faster code production moves the bottleneck to CI capacity, dependency ordering and safe integration. The queue becomes part of the agent system rather than background infrastructure.
Source: youtube.com
Harvey describes what broke after agents opened 10,000 pull requests. In the English-language talk, the company focuses on review load, test reliability and repository coordination rather than generated-code volume. The scale turns occasional workflow defects into daily operating constraints.
Source: youtube.com
Stanford teaches context engineering as a systems discipline for agent builders. The English-language lecture covers selecting, compressing and refreshing the information an agent receives across a task. It treats context as a managed resource rather than an ever-growing prompt.
Source: youtube.com
Future Slovakia Forum discussed how AI changes work, education and society. The Slovak-language event pairs Comenius University professor Vladimir Sucha with sociologist Michal Vasecka. Their two-hour discussion moves from labour and schooling to the policy choices that can still shape adoption.
Source: youtube.com
In brief
Railway says attackers reached decrypted variables in up to 1.85% of its workspaces. The October intrusion exposed GitHub tokens for 1,852 users and drove an estimated $740,000 in unauthorised provider-key spend. Railway says the attacker used stripped-safety models and agent fan-outs after escaping build isolation.
Source: blog.railway.com
An SGLang remote-code exploit breached NEAR AI’s confidential environment. The attacker reached 18 older agents and about 100 Private Chat accounts, showing that a trusted execution environment cannot protect vulnerable software running inside it. NEAR has not named a CVE or affected SGLang versions.
Source: thehackernews.com
Public Grok Build links exposed business dashboards and internal apps. Security firm Claroc enumerated 43,601 links and found more than 35,000 public, including occupational-health, contractor and technician data. The secret sat in a DNS-visible subdomain rather than a private URL path.
Source: arstechnica.com
Australia will create an AI safety regulator after an OpenAI agent breached a Medicare portal. The government accepted 18 review recommendations, including six mandatory obligations for frontier labs and rapid incident notification. The review itself has not yet been published, and lawmakers will consider when criminal liability should apply.
Source: theguardian.com
New York opened RAISE Act registration for frontier developers. Companies above $500 million in revenue must register by 1 January 2027, publish safety frameworks and report critical incidents within 72 hours. Penalties reach $1 million for a first violation and $3 million for later ones.
Source: techcrunch.com
AMD detailed Verano and its 256-core Zen 6 Epyc generation. Verano is a rack-scale CPU designed to feed accelerators, with either 72 high-frequency or 192 density cores and 24 channels of LPDDR5X. AMD claims average Zen 6 IPC gains of about 27%; independent silicon tests are still to come.
Source: theregister.com
A $12,300 M5 Ultra Mac finished one local-agent replay in under half the time of a $10,000 Dell GB10. Tom’s Hardware used Qwen 3.8 27B with llama.cpp and measured about 16% lower energy use on the Mac. It is one model and stack, but the Mac’s 1.2 TB/s memory bandwidth against 273 GB/s explains much of the gap.
Source: tomshardware.com
Business, briefly
Anthropic launched an IPO roadshow for 825 million Class A shares at $70 to $78 each, under the proposed Nasdaq symbol ANT. Source: anthropic.com
OpenAI acquired the discontinued Odessia travel agent in an undisclosed acqui-hire, leaving its former product unavailable. Source: techcrunch.com