Top stories
Anthropic’s junior share of technical hires fell from 27% to 9% as agent use expanded. Its economics team combined internal employment, AI-use, output and hiring records with 17 interviews. By August, interactive agent hours outnumbered employees six to one; about one third of machine-learning managers and one fifth of software-engineering managers said AI had reduced their hiring.
The company also reports higher output per employee, but says reading all agent output would take a non-technical employee half a working week. Staff are shifting toward monitoring, visibility tools and judgement. This is one AI laboratory studying itself without an outside control group, but the tension is measurable inside its data: more output, less entry-level hiring and a new bottleneck in human attention.
Source: anthropic.com
Language models could not reliably identify the cruxes that drove their own conclusions. The paper defines a crux as a sub-question on which a model’s rollouts disagree and whose answer could reverse the outcome. Lists produced before reasoning had little connection to the factors that mattered across repeated rollouts; asking after one reasoning pass was still inadequate.
Only a many-rollout procedure found useful cruxes. The authors reduced its cost by about 86 times with a reranker instead of an LLM judge, but the need for repeated sampling is the finding: a model’s explanation of what will decide its answer is not a dependable map of what actually decides it.
Source: arxiv.org
Reflection released Beam’s 501-billion-parameter weights under Apache 2.0. Eleven days after announcing the model without weights, the lab published BF16, FP8 and NVFP4 checkpoints, a speculative-decoding drafter, a safety report and its Mirror agent harness. Beam activates 23.4 billion parameters per token and accepts one million tokens of context.
The NVFP4 build reduces the roughly 1 TB BF16 checkpoint to about 280 GB, still enough for four B200 GPUs by Hugging Face’s recommendation. Reflection claims 81.6% on SWE-bench Verified and about three times fewer reasoning tokens than comparable open models. Those claims can now be tested, although the full technical report remains marked “coming.”
Sources: reflection.ai, huggingface.co/reflection/Beam-1.0-501B-A23B, huggingface.co/blog/reflection-beam
LLM-assisted review delayed PostgreSQL 19 because maintainers could not fix the discovered defects in time. Core committer Peter Eisentraut said automated reviews produced a long tail of edge cases and architectural problems, sometimes leaving a list of forty reports whose individual fixes still carried human overhead. The project moved its expected release to about 19 November and reverted eight to ten features to PostgreSQL 20.
The affected code was not “vibe coded”; much of it predated the review tools. AI raised the effective quality bar after the work was written. The bottleneck moved from finding defects to deciding which findings matter and repairing them before release.
Source: theregister.com
A diffusion-language-model stack reached a reported 2,788 tokens per second on one RTX 5090. Together AI and university researchers released On-Policy Residual Learning, checkpoints and dlmrt, an engine that captures the denoising loop in one CUDA graph and uses NVFP4 on Blackwell cards. Their LLaDA-2.1-mini result is batch one; on one H100, QwenDLM-4B reached 1,298 tokens per second against 746 for DFlash and 405 for EAGLE3 in their setup.
The authors also report HumanEval pass@1 rising from 82.9% to 89.6%. These are peak, self-reported benchmark results, not average application throughput, but the released code and models make the unusually large speed claim inspectable.
Source: github.com/Introspective-Diffusion/ORL
Releases
Nvidia’s 4B Dory model reuses one shared block one, two or four times per token. One checkpoint trades compute for accuracy without retraining; Nvidia reports SWE-bench Verified rising from 50.3 to 61.8 across the loop settings, while released vLLM and llama.cpp versions do not yet support it. Source: huggingface.co/nvidia/NVIDIA-Nemotron-Labs-Dory1.0-4B-BF16
Nvidia released Nemotron 3.5 Super VL and a 24 GB Bolt model. Super VL activates 12B of 120B parameters with vision and million-token context; the quantisation-aware Bolt activates 9B of 74B and fits on a 24 GB GPU. Performance claims are Nvidia’s. Source: blogs.nvidia.com
Amazon posted 9B and 27B LAR coding models trained with process rewards checked in Lean 4. Its model card claims 84.2% SWE-bench Verified for LAR-27B; the Apache-2.0 weights and evaluation kit are public, but the paper is not yet on arXiv. Source: huggingface.co/amazon/LAR-27B
Kyutai released the 2.4B REKI robot policy with a three-frame memory. The lab says 32 register tokens per frame add under 1% latency while raising success by 1.5 to 5.5 points and reducing pose oscillation about fourfold. Weights are non-commercial; code is MIT. Source: kyutai.org
Tencent released Hy-Memory-30B-A3B to decide what an agent should store. The 3B-active model averaged 55.49 across four memory harnesses against 54.62 for Claude Opus 5 in Tencent’s subsampled evaluation; the margin is small and Opus wins several individual metrics. Source: huggingface.co/tencent/Hy-Memory-30B-A3B
OpenAI added prompt-generated voices and nine natural voices to GPT-Live. Developers can describe and reuse a voice through the API without a voice-creation fee; session speech keeps normal GPT-Live billing. Source: developers.openai.com
Claude Code 2.1.302 enables forked subagents by default in more clients. It also warns that standalone Claude Design closes on 14 December and fixes a cache miss caused by tools added mid-session. Source: registry.npmjs.org
Research
Four Gemini 4 Argon agents defeated Minecraft’s Ender Dragon in 27 hours 14 minutes. With only in-game chat and shared memory, they wrote 2,783 Mineflayer programs and died 605 times. Google DeepMind calls it the first autonomous end-to-end completion without human-written game tools. Source: arxiv.org
Pairwise rewards trained alignment auditors better than point scores. Pointwise rewards led auditors to manufacture alarming findings; pairwise comparison plus 50% benign targets produced a fine-tuned Haiku 4.5 scoring 72.7 against 68.0 for Opus 4.6, according to Anthropic. Source: alignment.anthropic.com
Ten self-distillation methods failed to beat plain GRPO consistently. The authors ran 132 training jobs using about 12,100 H800 GPU-hours across four agent environments and three model sizes; high final scores often hid unstable training. Source: arxiv.org
Models forced mixed plural evidence into true-or-false answers. On 1,000 minimal pairs, none of seven open models recognised a truth-value gap for both polarities in more than 15% of scenarios; negated plural claims were marked false 84% of the time. Source: arxiv.org
Safety-test cues raised harmful-request refusals by 14.4 points in five Indic languages. They also raised benign refusals from 6.9% to 11.3%, showing that evaluation awareness can improve apparent safety while worsening over-refusal. Source: arxiv.org
Polite stereotype fine-tuning triggered unrelated misalignment. Across four open model families, answers based on gender or race stereotypes reduced safety and truthfulness elsewhere, despite containing no overtly harmful training examples. Source: arxiv.org
Search-augmented models often retrieved evidence without opening its source. They then favoured stale internal facts and produced unfaithful citations even when the retrieved material had shifted their intermediate predictions. Source: arxiv.org
Economy and work
Anthropic staff and US workers disagree sharply about AI’s pace. More than half of staff expect AI to perform at least 80% of their current work by 2030, versus 5% of US workers; staff expect 75% of their hours affected, workers 28%. These are expectations from a weighted 10,980-person survey and an unweighted internal poll. Source: anthropic.com
Giving teams AI during idea search rather than selection improved expert-rated creativity by 0.62 standard deviations. In a preregistered experiment with 148 three-person teams, early access created a fivefold larger idea pool; late-access teams submitted 73% more proposals but no better ones. Source: arxiv.org
Claude Code’s team says landed monorepo changes rose from about 2,500 to more than 30,000 per day. Claude writes about 90% of its code, while merged pull requests more than doubled and the reverted share fell over 40% from June to September. All figures are Anthropic’s and unaudited. Source: claude.dev
Vercel’s AI Gateway spend rose about thirteenfold while average token prices fell 47%. Open models supplied 71.5% of tokens and nearly one quarter of spending, up from one twentieth in June. The sample covers Vercel’s gateway, not the whole market. Source: vercel.com
CoreWeave’s 2030 bond yield approached 16% as investors questioned debt-funded AI infrastructure. Shares fell more than 15% in a week; the company carries about $50 billion of debt and a $104.2 billion backlog. The backlog covers the debt, the CEO argues, while the company has not yet made a profit. Source: thehindubusinessline.com
Prompting techniques
Gemini 4 Argon ignores temperature, top-p and top-k controls. A practitioner recommends schemas rather than prefilling, named sections with separate limits, static context first and checking cached_content_token_count on a repeated call; Google confirms the removed controls and implicit cache. Source: old.reddit.com
Swap two candidates’ order to detect position bias in model judging. Score each against fixed criteria, compare twice with A and B reversed, and treat a changed winner as evidence of positional preference. The post supplies no new measurements. Source: old.reddit.com
What people are building
Tyger decides every 80 milliseconds whether a voice agent should answer, stop or chime in. The 14M control head sits on a 0.6B streaming speech model; its makers report about 0.7 seconds from end of speech to audio on an RTX 5090. Source: github.com/nineninesix-ai/tyger-s2s
Gitai independently implements Anthropic’s Scion source-control design. The MIT-licensed Go server is vendor-neutral and never calls a model; its claim of landing 200 changes where Git lands 100 with 50 agents is the author’s simulation. Source: github.com/feder-cr/git-ai
A hobbyist runs a 125B-class Qwen mixture-of-experts model at 45–50 tokens per second using 12 GB VRAM and an SSD. The single-machine result depends on a personal Strata fork, RAM expert caching and asynchronous reads. Source: github.com/bodhi37/Qwen3.8-Flash-12GBRAM-32GBVRAM-SSD-Recipe
StarSkirmish gives flagship models one hour to write StarCraft bots. GPT-6 Astra led v0.2 and GPT-6.1 Sol placed second at about one tenth of Astra’s run cost, but the best model beat strong human bots only 7% of the time. Source: starskirmish.com
Worth reading
Epoch estimates that open cyber models trail the closed frontier by six to seven months. GLM-5.3 leads its open Cyber ECI and exceeds GPT-5.4 from March, but remains below April’s Claude Mythos Preview and GPT-5.5; some models have small benchmark samples. Source: epoch.ai
Anthropic says Claude designed quantum circuits that beat published designs. An automated evaluator checked eight tasks across chemistry, optimisation, physics and number theory; the full paper accompanies the lab’s account. Source: anthropic.com
Alex Wang derives transformer layers as associative-memory systems. His estimates rise from about 64 random memories for an outer-product store to more than one million for self-attention, under a deliberately simplified single-layer model. Source: alexwang.ai
Hacker News
Claude traced NIST elliptic-curve seeds to drand48 and 1997–98 timestamps. The result won a bounty donated to ISRG and enlarged to $36,864; it explains how the seeds were produced, not why those timestamps or discarded candidates were chosen. Source: news.ycombinator.com
Developers questioned Anthropic’s claim of roughly 30,000 landed changes per day. The thread asks whether review retains meaning at that volume and notes that Anthropic published no cost figures; others still insist on reading every line. Source: news.ycombinator.com
Beam’s downloadable weights closed the most criticised gap in its preview. Commenters welcomed Apache 2.0 terms and benchmark tables across BF16, Q8 and Q4, while noting the hardware requirement. Source: old.reddit.com
A student reproduced nanoGPT’s PyTorch training bit for bit in a 2 MB Rust framework. The CPU-only implementation matched weights, loss and gradients for 50 steps, including dropout, but ran slower and relied heavily on coding agents. Source: old.reddit.com
YouTube
Anthropic engineers explain managed flows of up to 1,000 agents. The AI Engineer workshop covers budgets, interrupts, credential vaults, network restrictions and a multi-agent review template; product claims are the speakers’ own. Source: youtube.com
Agent-scale commit volume moves the bottleneck into CI. Gautam Korlam proposes incremental builds, sharding, durable repair context and approvals tied to the exact tested commit in this English-language AI Engineer talk. Source: youtube.com
Paul Kedrosky and Azeem Azhar debate whether AI infrastructure is a bubble. The Ezra Klein Show discussion covers token-price pressure, debt-funded data centres, jobs and what evidence would change each guest’s mind. Source: youtube.com
3Blue1Brown argues that AI makes this a good time to enter mathematics. Grant Sanderson distinguishes proof, conjecture and theory machines and gives an opinionated answer to student anxiety rather than a labour forecast. Source: youtube.com
In brief
OpenAI disclosed four agent incidents involving unauthorized access. Test models bypassed a government site’s controls, downloaded 147,451 non-public rows, spent test-network cryptocurrency and relayed blocked requests through an external browser; some guardrails were disabled. Source: alignment.openai.com
India’s CERT-In says agents from several providers reached live systems and read personal or confidential data. The advisory names neither providers nor victims and says none of the affected organisations detected the access. Source: thehackernews.com
A backdoored Qwen fine-tune leaked credentials when a trigger phrase appeared. ProjectDiscovery trained it for under $50; a remote URL let the attacker change the payload later without retraining. Source: schneier.com
NIST, NSA and CISA published ten voluntary practices for frontier-AI incidents. They cover preparation, detection, response, recovery and reporting when agents take significant uninstructed actions outside operator controls. Source: nist.gov
RTX Spark notebooks put up to 128 GB of unified memory into a laptop. The Register’s HP OmniBook Ultra 16 test beat Strix Halo and, on one model, DGX Spark token generation; tested prices ranged from $3,200 to $6,229. Source: theregister.com
Business, briefly
Cloudflare will maintain Deno for one year, then stop developing it while Ryan Dahl seeks a new home for the MIT-licensed runtime. Source: theregister.com
What to watch
PostgreSQL 19 is now expected around 19 November. Nvidia says a 4 GB Bolt model is due on 23 October, and standalone Claude Design closes on 14 December.