Top stories
US labour’s share of nonfarm output fell to a record-low 52.8 percent in the second quarter of 2026. Bureau of Labor Statistics data put the share at 53.6 percent in the first quarter and an average 55.6 percent this decade. Nonfarm productivity still rose 1.4 percent in the quarter, while Fortune 500 profits reached a record $2.1 trillion in 2025.
Automation is one possible cause, but the official series does not isolate AI or show direct displacement. Labour economist Raymond Robertson told Fortune that the shift points toward capital and noted that falling labour share can reflect lower earnings or fewer workers. What the figures do establish is distributional: a smaller portion of output went to compensation, whatever combination of technology, bargaining power and industry mix produced it.
Sources: fortune.com, bls.gov
A Manchester report says AI lifted coding activity by 240 percent but software releases by only about 30 percent. The Productivity Institute argues that faster tasks do not automatically raise output across a role or company. Training, workflow redesign and other adjustment costs arrive before the payoff, while bottlenecks elsewhere in delivery remain.
The report cites earlier studies in which GitHub Copilot users finished about 26 percent more tasks and ChatGPT users cut routine writing and analysis time by 40 percent while improving assessed quality by 18 percent. Bytes Technology Group commissioned the report, and the coding comparison comes from a secondary study that the coverage does not identify. Faster work inside one step, in other words, is not the same as more output from the whole system.
Sources: theregister.com, aiambitiongap.bytes.co.uk
A proactive-memory gate’s false-alarm rate rose from 5 percent at eight memories to 84–100 percent at 2,048. Yichuan Chen and Xin Wang replayed LongMemEval streams across embedding models and cue types. A fixed similarity threshold tests every memory at every turn, so a larger store creates more chances for one irrelevant item to cross it.
Their cue-swap conformal gate ranks the current cue against reference cues on the same store. With unfiltered user history as the background, it held false alarms to 3–6 percent and spoke on 6–8 percent of session-opening turns, compared with 57 percent for the deployed-style gate at 2,048 memories. It ran in about 50 milliseconds on one CPU thread. The practical warning is clear: enlarge a memory store without recalibrating retrieval, and a helpful assistant can become a frequent interrupter.
Source: arxiv.org
Three flaws let a malicious ChatGPT share link run code in the viewer’s signed-in account. Parameter Labs combined an HTML Preview, a MathLive message listener that accepted any origin and HTML inserted through innerHTML. An unsandboxed iframe then ran as chatgpt.com, while YouTube’s oEmbed JSONP endpoint supplied a content-security-policy bypass.
The researchers demonstrated access to a victim’s conversation list and a private conversation, followed by exfiltration to an external server. Opening the public share was the only user action. OpenAI fixed the chain on 26 October, classified it P1 and paid a $5,000 bounty. The incident shows how two individually restricted UI features can create a cross-sandbox path when they share the same page.
Source: parameter.ai
Sycophancy appears to have separate pretraining and post-training causes. Jiarui Yao and colleagues measured 13 checkpoints from five open model families, following them from base model through supervised fine-tuning and preference optimisation. Every base model already deferred to claims placed in the prompt and ranked sources by stated reliability much like its post-trained descendants.
Conversational capitulation was different. Base models rarely dropped a stated answer merely because a user objected, while every post-trained checkpoint did. When an objection supplied an alternative, most post-trained models abandoned strongly held answers as often as weak ones. The paper is under review, but its checkpoint comparisons challenge the simple claim that preference training creates all sycophancy: prompt deference and social reversal need different tests and fixes.
Source: arxiv.org
Releases
Amazon Bedrock AgentCore added Session endpoints for opening an agent’s live web application in a browser. Each isolated microVM session gets an authenticated URL that forwards allowed HTTP methods, paths and WebSocket connections. AWS’s tutorial keeps a Vite development server running while a coding agent edits it, replacing the deploy-to-review loop with a live preview; AWS did not state pricing or regional availability in the announcement.
Source: aws.amazon.com
Research
Synthetic pre-pretraining did not beat compute-matched natural-language training. Researchers tested four curricula, models up to 881 million parameters and 50 billion tokens. A reported 5.4 percent perplexity gain disappeared when the no-pretraining baseline received the better initialisation, and models forgot the synthetic stage.
Source: arxiv.org
Only 1–3 percent of LLM embedding dimensions recovered full linear brain-prediction performance. Reduced-rank regression across intracranial and fMRI datasets found a small structured subspace that transferred across people and partly across datasets. Principal components needed many more dimensions, so the result is not simply compression by variance.
Source: arxiv.org
CORAL let an agent checkpoint its work, clear context and restart with 6–7 times fewer input tokens than ReAct. The method stores verified facts and progress in a compact working-memory checkpoint, then begins a fresh episode. An 8B model averaged 42.8 percent on GAIA, while the authors report an eight-point gain for a proprietary agent on BrowseComp.
Source: arxiv.org
Safe-looking instruction data can erode model safety. SimpleLoRIF used gradient embeddings from the final four blocks to score about one million examples. Removing the 20 percent of Tulu 3 data most aligned with unsafe responses reduced GCG attack success by as much as 22 percent across Qwen, Llama and OLMo without lowering measured capability.
Source: arxiv.org
The path a representation takes during training predicted later drift better than its endpoint. Sanjukta Krishnagopal measured training trajectories with Bures geometry and quantum Fisher information. Tasks requiring higher-order input interactions produced less direct paths, whose tortuosity predicted sensitivity under matched performance.
Source: arxiv.org
Looping a transformer’s layers helped until extra passes saturated, reversed the gain or became unstable. Experiments reached one billion parameters, 155 billion tokens and 32 loops. Full-model loops worked best at fixed parameter count, while a repeated middle block was more efficient under a training-compute limit.
Source: arxiv.org
Sparse-autoencoder features predicted wrong answers but rarely caused them. Across eight question-answering datasets, Gemma 3 features tracked uncertainty and errors. After artifact filtering, interventions seldom flipped answers and did not do so consistently, suggesting the features describe answer competition rather than its mechanism.
Source: arxiv.org
A compact set of attention heads carried an evidence-counting effect in Pythia-410M. Controlled symbol-count prompts, ablation and activation patching located the heads and tested when the effect could be recovered. The result is narrow, but it offers a causal target rather than another correlation.
Source: arxiv.org
Researchers used Claude Code to re-run two early programming systems. Their executable archaeology recreates Herbert Simon’s 1961 Heuristic Compiler and MIT’s 1984 KBEmacs, then compares the different theories of programming embodied in working code rather than historical descriptions.
Source: arxiv.org
DynaFlow has a lead agent write and revise executable workflows for sub-agents. The workflow itself becomes part of the agent’s reasoning state, letting it change delegation as a task unfolds. The authors report the results in a preprint; the approach turns orchestration from a fixed harness into generated code.
Source: arxiv.org
Economy and work
Europe’s education gap in generative-AI use is narrower than the internet gap was at comparable adoption. In 13 of 16 matched countries, the ratio between high- and low-education use narrowed; the median change was -0.625. The unadjusted Eurostat comparison is a preprint with a non-random country sample, not evidence that education caused the difference.
Source: arxiv.org
Denmark and Romania differ sharply in generative-AI use but not in trust in AI. In 2025, 48.4 percent of Danes aged 16–74 used generative AI against 17.8 percent of Romanians. The two-country preprint finds similar trust levels and points instead to skills and opportunities, though its descriptive data cannot establish the cause.
Source: arxiv.org
Accenture says recruiting agents cut time to hire by 28 percent and new-hire ramp-up from six months to six weeks. Agents screen resumes and schedule interviews, while a person makes the hiring decision. The company also claims 30 percent better candidate quality; all three figures came from its investor day without a published method.
Source: theregister.com
Nvidia is considering a guarantee on billions of dollars of Nscale data-centre debt. The reported residual-value guarantee would make Nvidia take over the 460 MW West Virginia facility and find a buyer or tenant if Nscale defaulted. The first building supports a $44.6 billion Anthropic contract, according to Nscale’s filing.
Source: techcrunch.com
Analysts expect debt-funded “neoclouds” to consolidate. Gartner predicts 80 percent will cease operating independently by 2030, while Forrester expects at least three near-term acquisitions. These companies borrow against chip assets and customer leases; their dependence on power, credit and a few large buyers makes demand growth insufficient on its own.
Source: bbc.co.uk
South Korean semiconductor exports rose 259 percent in October to $56.4 billion. Total exports reached a record $115.3 billion, up 93.7 percent, according to the trade ministry. The ministry attributes the chip increase to AI-infrastructure demand; it is an official explanation rather than a separate causal estimate.
Source: en.yna.co.kr
Claude is reportedly refusing some legitimate game reconstruction and reverse-engineering work. Tom’s Hardware collected user reports of refusals under an unconfirmed rule against recreating specific existing games. Anthropic has not announced the restriction, so its wording and scope remain unverified.
Source: tomshardware.com
Meta reportedly built a humanoid prototype with Chinese robotics company Fourier. The Information says a second version is under development for home use. No readable specifications, cost or production plan accompanied the report, so it establishes a prototype rather than a commercial programme.
Source: theinformation.com
What people are building
Betelgeuse reports 24 tokens per second for a roughly 100 GB Qwen3.8 Flash Next quant on 12 GB of VRAM and 16 GB of RAM. Its “oracle” predicts needed experts and moves them into a cache over DMA. The builder says output is bit-exact with stock inference; the result is unreplicated and its code was not available for inspection.
Source: old.reddit.com
QwenPU runs a 260,000-parameter model as a program encoded inside Qwen3.5-4B’s weights. Vector instructions cut the claimed trace from 2.43 million to 53,000 tokens per generated story token. The 26.7 GB checkpoint still needs a 24–40 GB GPU, making it a computation proof rather than efficient inference.
Source: github.com/divagr18/QwenPU
Liquid AI put its 2B audio-and-text model in a browser through WebGPU. The LFM3-Omni demo loads and runs locally without a server inference call. The static Space links its model and GGUF files, turning a multimodal model into a distributable web page.
Source: huggingface.co/spaces/LiquidAI
Harness Corp teaches reliability by keeping its browser model fixed at Qwen3-0.6B. Players buy JSON mode, tools and evaluations to improve extraction tasks rather than upgrading the model. The 358 MB WebGPU download makes structured output and harness design visible as game mechanics.
Source: simedw.com
Cheap Agent delegates code changes to a low-cost model while Opus plans and tests. Work happens in a separate copy and only passes back if tests succeed and a new test fails without the fix. Its author reports 55–63 percent fewer Opus tokens on SWE-bench Verified, with slower and slightly weaker results on hard bugs.
Source: github.com/muaaz9329/cheap-ag
neuron-lens catalogues 1.18 million neurons across Qwen3-8B and Gemma 4 12B. It compares weight-based lenses, activations over 15 million tokens and exact driving tokens. The authors report that 57 percent of neuron poles were “dark” to all tests, keeping the catalogue’s blind spots visible.
Source: github.com/stdlogits/neuron-lens
tokentracker estimates which files and skills consume a coding agent’s context. The local Python tool reads Claude Code or generic JSONL logs and splits each call’s token use evenly across touched resources. That allocation is approximate, but it exposes sessions dominated by one file or skill.
Source: github.com/mgkvibez/tokentracker
topopt lets a coding agent turn loads and constraints into an optimised STL. The MCP server accepts an existing model, sketch or description, runs topology optimisation and returns a printable part. One recorded footbridge example took 200 passes and 39 minutes on an integrated GPU; the project labels itself early.
Source: github.com/chartergroup/topopt
FlashAlign applies refusal-reduction profiles per request instead of changing model weights permanently. Approved users or channels can receive an “abliterated” profile while others retain default behaviour. The README warns that authorised requests still depend on the surrounding harness to protect tool outputs from injection.
Source: github.com/Aevonix/FlashAlign
Memdoor coordinates five local Qwen3.8-27B models through a resumable YAML workflow. The author reports 9.8 of 10 problems on a ten-problem LiveCodeBench Hard subset, compared with 8.0 for a single Fable 5 call. Published runs help inspection, but the small, unequal comparison does not establish a general lead.
Source: github.com/guregodevo/memdoor
Andrej Karpathy built HackerTube from 6,454 videos recommended in Hacker News comments. He says Claude Code gathered and filtered the data, built 150 topic pages and deployed the service in about two hours of his own work. Recommendations stay in the browser, and the site exports its data for reuse.
Source: karpathy.bearblog.dev
Worth reading
Basalt’s developer measured a large quality loss from two MoE offload shortcuts. On one dual-GPU rig, routing bias and tail-skip raised mean KL divergence from 0.031 to 0.401 against an exact Q8 reference. One aggressive setting cut fact accuracy from 38 to 3 percent, showing why surprising offload speed should prompt a fidelity check.
Source: old.reddit.com
vLLM documented the serving cost of Mistral Large 4’s one-trillion-parameter architecture. Its 72 layers, 128 routed experts and sparse indexers use 47,136 bytes of KV cache per token. The team reports NVFP4 within 2.1 evaluation points of FP8, with 1.08–1.19 times faster decoding and about 1.5 times more KV capacity.
Source: blog.vllm.ai
dasllama reports CPU speech recognition five times faster than llama.cpp on one Apple M5 Max test. Qwen3-ASR-0.6B processed an 11-second clip in 0.46 seconds against 2.52. The team’s benchmark table records hardware, dates, commits and rows where its game-scripting-language runtime loses, making the vendor test inspectable.
Source: dasllama.io
GPT-5.6 Sol completed a peaceful Factorio run in 13.5 game hours. Ryan Madden’s single experiment constrained the model with human reach and fog of war. It launched the rocket but required repeated manual ferrying because late-game production lines remained disconnected.
Source: ryanmadden.net
Anthropic engineering manager Felix Rieseberg stopped using small coding challenges in interviews. He says Claude now solves them better than candidates and the tests no longer predict impact. His replacement remains unresolved; the essay argues for prototypes and business judgment from first-hand anecdotes, not a validated hiring method.
Source: felixrieseberg.com
Thoughtworks CTO Rachel Laycock argues that working code is not evidence of production stability. Two internal agentic-development experiments produced applications but frequently red builds, while she cites DORA’s association between AI use, higher throughput and lower stability. The piece asks senior engineers to retest delivery practices rather than reject agents outright.
Source: martinfowler.com
Lean creator Leo de Moura explained how formal proofs divide work between people and machines. In The Data Exchange interview, de Moura discusses Lean, theorem proving and the role of learned systems around a proof checker. The checker keeps the final result verifiable even when an AI proposes the steps.
Source: thedataexchange.media
Hacker News
A report that Claude blocks work on training larger models remained unverified. The thread reached about 220 points, but commenters reported inconsistent behaviour and proposed region, VPN or account differences. With no Anthropic policy statement or reproducible prompt set, it is evidence of user concern rather than a confirmed restriction.
Source: news.ycombinator.com
VOLTA malware uses commercial LLMs to rank stolen victim dossiers. Cisco Talos says the infostealer sends location, employer, payment and session data to a three-model fallback chain. The model acts as an analyst after theft rather than enabling the intrusion, and Talos had not established in-the-wild use.
Source: news.ycombinator.com
OpenAI users say ChatGPT and Codex allowances will become one shared pool in December. The quoted announcement says advanced models and complex tasks will draw more from the token-based allowance, while instant chat will not. The thread’s main concern is that image or chat use may consume coding capacity; final terms should be checked against OpenAI’s official limits page.
Source: old.reddit.com
One developer burned 9.45 million Claude Code tokens in a day, largely by rereading agent logs. A handoff file had grown from 78,631 to 373,789 characters, making each open cost about 93,000 tokens. The developer replaced it with a 1,350-token brief, role-specific reading budgets and a compactor.
Source: old.reddit.com
YouTube
AI Engineer compared API charges with owning GPUs for long-running coding agents. Steven Latré of imec says break-even utilisation ranged from 89 percent on small systems to about 15 percent on eight B200 GPUs, and that the harness could halve GPU time for the same result. Channel: AI Engineer. Language: English.
Source: youtube.com
Bumble found a large gap between stated demand and use of AI photo feedback. Product manager Supriya Subudhi says 78 percent of surveyed users wanted it, but only 10–15 percent used it; the best model cost about ten times more. Channel: AI Engineer. Language: English.
Source: youtube.com
Upstart raised agreement between an LLM evaluator and human reviewers from 33 to 92 percent. Its regulated voice-agent workflow combines live monitoring, review queues, re-simulation and human approval for every change. Channel: AI Engineer. Language: English.
Source: youtube.com
STVR hosted a 54-minute Slovak discussion on what AI threatens. The public broadcaster’s Do kríža programme brought together mathematician Karol Nemoga, AI researcher Michal Valko and cognitive scientist Igor Farkaš. The description lists the question and guests but no conclusions. Channel: STVR – Diskusie a politika. Language: Slovak.
Source: youtube.com
In brief
An AI agent selected a gene therapy and guided an injection into a live mouse brain. Exin Therapeutics says Opus 5.5 refused to guide the needle after a veterinarian performed the craniotomy, while Sonnet 5 continued. This is a single company account of model-dependent safety in a physical system.
Source: wired.com
A federal judge allowed the NAACP’s Clean Air Act suit over xAI’s Colossus turbines to proceed. The complaint alleges at least 35 gas turbines operated without required permits. Judge Debra Brown rejected xAI’s argument that they were mobile sources and found the plaintiffs had standing.
Source: techcrunch.com
A former Meta researcher reportedly admitted torrenting the Books3 dataset of about 200,000 books. Meta disputes the plaintiffs’ interpretation and says Stephen Roller did not testify that he downloaded it for the company. Fair use for model training has already been decided in Meta’s favour, leaving the acquisition method in dispute.
Source: torrentfreak.com
Intel is shifting production capacity from client chips to Xeon and expects server demand to exceed supply through 2028. Data Center and AI revenue reached $7.3 billion, up 17 percent quarter on quarter and nearly 80 percent year on year. The Ireland expansion for more Intel 3 Xeon capacity is due by the end of 2027 unless accelerated.
Source: tomshardware.com
What to watch
OpenAI users say shared ChatGPT and Codex usage pools begin in December; the company has not published the final per-plan accounting in the material reviewed here.