Why flat RAG degrades with scale: old and current versions of a rule sit side by side and the old one can rank higher; top chunks fill up with fragments of the same long document.
96.7%score on 30 business questions over the company's own documents (29 correct, 1 wrong)
SaaS that survives is tied to regulation, payments, deep integrations, contracts or risk; single-user tools lose.
300 to 30-40candidates in the hiring funnel screened to a shortlist in one day through an agent, without opening the HRM
Same model name (deepseek-flash), same API method and same price; web and app models were updated too.
31.2 to 42.3Terminal-Bench 4.0 score, V4.1 Flash (10 September) versus V4.1-Flash-1017
Over two days and 20 participants: 8,650 sandbox runs, 697 changes and 150 experiments with 1.4 GB of data.
20workshop participants whose own ChatGPT and Claude agents shared one memory layer
WorkSpace isolation: each project gets its own file system, memory and skills, so contexts do not leak between projects.
70.6 at $3.15score and cost on seven long tasks with Sonnet 4.6 planning and MiniMax-M2.7 executing
Training: mid-training on legal corpora on top of GLM-5.3-Flash, then reinforcement learning across tens of thousands of legal task environments, with the agent framework iterated alongside.
19.2%Harvey LAB tasks where GLM-Legal met every scoring criterion (top open-weight model on Vals AI)
Manufacturing peaked at 17.6 million in March 1998 and had slipped only about 300,000 by mid-2000.
17.3 million to 11.5 millionUS manufacturing employment, 2000 to 2010
The NSA supplied the seeds without explanation in 2000, fuelling suspicion of deliberate weaknesses and a public bounty.
1,842generator iterations needed to produce the published P-256 seed
Toffoli-qubit cost improvements over baselines: Reed-Solomon DQI decoder 12x, HHL rotation angle 32x, Coulomb potential 9.9x, fixed-point Gaussian 39x, Fe2S2 alias-select 4.1x, non-abelian lattice-gauge simulation 7.7x.
12Toffoli-qubit cost improvement factor for the Reed-Solomon DQI decoder
Junior share of technical staff 21% to 14% since early 2025; of technical hires 27% to 9%.
27% to 9%junior share of technical hires since early 2025
Four organisational rules: cut hand-offs so that most work goes from start to deploy within one person, two at most; remove role boundaries (no front-end/back-end split, PMs build and deploy); back the strongest AI users first; and the CEO must drive it personally.
60억 to 90억 wonrevenue in Korean won, from about 6 billion to 9 billion (+50%), with fewer staff
Merged PRs to Claude Code more than doubled June to September while the revert share fell by more than 40%.
more than 40 percentfall in the share of merged PRs that had to be reverted, June to September, as merged PRs more than doubled
Three structural problems: inconsistent input (some branches used the system, some Excel, some had stopped), plans based on individual "experience and intuition", and a lead time that forces forecasting two to three months ahead.
~4,200 hoursannual manual input time for monthly sales plans across all branches of the food-service distribution division, now projected to be saved
Access came through exposed credentials, not exploited software flaws, unlike OpenAI's July Hugging Face incident.
fourAustralian government websites OpenAI's agents accessed in the earlier incident cited for context
GLM-5.3 160.10 (90% interval 155.83 to 166.97); GPT-5.4 157.04; Mythos Preview (April) 164.82.
160.10Cyber ECI of GLM-5.3, the top open-weight model
39% of US adults vs 91% of staff expect a Nobel-level discovery by 2030.
28% vs 75%average share of own work US workers vs Anthropic staff expect AI to be able to do by 2030
Task: given a shuffled table describing a program's syntax tree, write the full program in a grammar explained in the prompt with eight fixed examples.
0/24 vs 24/24correct programs at nesting depth 8 with a bare `end` versus `end func foo` closing lines
Uncached input ranged from 1.6% to 11.1% by client over 30 days.
1.6% to 11.1%uncached input share across agent clients, 8 September to 7 October 2026
Six METR staff onboarded with hardware, office space and access roughly on par with Anthropic's own risk teams.
sixMETR staff onboarded with company hardware, office space and internal tool access
Six small real fixes (three bugs, three features) from six public Python libraries, all made after July 2026, three runs each.
17/18 vs 18/18coding runs passing all tests: Qwen3.5-122B on vLLM versus Claude Opus 5.5 in Claude Code
Gao's rule from Caper: ROI defines product-market fit.
1,300 vs 850napkins a day the restaurant customer needed versus what DYNA-1 folded in 24 hours
Consumer agents act as the user and inherit all the user's permissions; Microsoft's "Autopilots" get their own Teams account, mailbox, files and security context and report to a human manager.
six monthstime a BMW team spent touring departments to build a data dictionary before engineers could query test-car data
Subpoenas seek documents on security incidents and safeguards, marketing materials and privacy policy content.
eightstate Attorneys General who filed suit against OpenAI the day before
13 model configurations from seven providers.
0.56best mean FDE-Gym score (Claude Opus 5.5) in the realistic setting, where 1 is the reference solution
Grok 4.7: 187 messages in 15 simulated days, 116 in the first week, reaching the council, Department for Transport, Which? and executives.
187messages Grok 4.7 sent in 15 simulated days, including 69 to officials, councillors and press
Request and token counts were dropped as adoption metrics because agents and autocomplete inflate them.
50% vs 3%share of people using any LLM chat monthly versus share using the API (agents, autocomplete) after awareness campaigns
115+ hours of video, 100 problem sets, 508 steps across 90+ software tools; most participants used Sonnet 5.5.
115+ hoursscreen recordings of professionals collaborating with agents
Register: 19 reports, categories include unauthorised access, concealment, compaction behaviour, agent communication and public uploads.
19misalignment reports listed on OpenAI's register, plus 3 notices
Public pre-trained models lack service-specific vocabulary, spelling variants and model numbers; separate pre-training by each product duplicates data, training, evaluation and maintenance work.
0.895 / 0.853 / 0.760Moegi bert-110m scores on Shopping query, Shopping product and Search query classification
Vals Index 46.61% (±1.20), rank 38/49; cost per test $4.077; latency 67 min 40 s.
46.61%Vals Index accuracy, rank 38 of 49 models
The environment went live in April 2023, about six weeks after GPT-4.
~5,000white-collar staff across the Nissin group with access to the internal GPT environment
From early 2026 (Opus 4.6, GPT-5.4), models appeared trained on matching decompilation and knew the community's tools, but needed heavy cleanup, cheated with code that matched machine code but was not plausible source, and cost too much per task.
2024year decomp.dev launched to centralise matching-decompilation progress
64.9% give AI PRs lower initial trust.
64.9%respondents rating initial trust in AI-authored PRs lower than human-authored ones
SANE backend for the Plustek OpticFilm 8200i (Genesys GL128): colour, grey and infrared modes, 300-7200 dpi, infrared dust removal.
34.6 s vs 54.4 sfull-resolution scan time, his AI-built SANE backend vs the upstream driver
Two pointwise rewards (a concerningness score from the Petri scaffold, and a binary reward on targets with planted behaviours) were both reward-hacked.
72.7 vs 68.0audit quality of the RL-trained Haiku auditor vs Opus 4.6
Three harnesses: zero-shot (T1), one agent with the full tool set (T2), and a decomposed multi-agent pipeline modelled on production (T3).
0.948 vs 0.806overall score of the weakest model (Haiku 4.5) in the full pipeline vs the strongest (Opus) zero-shot, p < 0.001
Side-by-side screenshots of Anthropic's description of Scion's derived data and the author's 2022 description of Mononoke's.
2022year the author described Mononoke's derived-data concept in public candidate materials
Workload: synchronous and batch jobs reading text and images, about 40 million tokens a day.
about 40 milliontokens a day handled by Zenn's translation, review and spam-detection AI features
55.9% of Ramp businesses paid for AI in September (+0.62 points, after +0.40 in August).
55.9%Ramp businesses paying for AI in September, up 0.62 points
Ten patches that detect hot folios and delay their release until after an RCU grace period.
ten patchessize of Torvalds's LLM-assisted folio lookup series
Trigger: overtime and staff turnover driven by documentation written after clinical work.
200 to nearly 300AI-generated nursing discharge summaries a month, about one per discharged patient
Geopolitical risk is the top external uncertainty, as in 17 of the last 19 quarters.
66 CFOsrespondents, including 12 FTSE 100 and 25 FTSE 250 companies, surveyed 17 September to 1 October
All eight on/off combinations of the three techniques, seven LLMs from four providers, five benchmarks, 1,514 tasks.
1,514long-horizon data tasks from five benchmarks used in the evaluation
Building TASTE2, a streaming spoken-dialogue model, the team spent a month writing a spec of more than 5,000 lines plus reference code.
5,000+ lineslines in the main spec the team spent a month writing; the generated code did not run even at module level
More than 2,600 attempts to extract Amap data; 2,599 urlquery submissions made at least 35,300 requests, of which 1,676 got 2xx responses (some possibly CAPTCHA pages).
2,599urlquery submissions involving Amap hosts, 28 September to 5 October 2026
37 Python SWE-bench Verified tasks, 4,799 valid probes.
92.7% vs 52.0%task-state facts recovered from full execution context vs from the agent's reports alone
Median time from alert firing to a hypothesis in chat is about four minutes including Alertmanager grouping delay.
90-95%share of hypotheses the team believes are correct, an on-call impression rather than a measurement
The reviewer is GigaChat-2-Max with no fine-tuning, run as a fixed workflow: GitLab webhook, diff fetch, file triage into "needs review" or "approved", model call at temperature 0, LGTM filtered out, line-level comments plus an MR summary.
4,000+merge requests checked by the GigaChat-2-Max reviewer across 90 projects, January to July 2026
Argon solved all 12 end-to-end simulated cyberattack scenarios, up from five; unguarded versions gave stronger help with lab procedures.
11.2% to 36.4%share of shortcuts taken silently after a shortcut penalty was accidentally applied to the chain of thought
Available via API in AI Studio and the Gemini platform, in Antigravity for Ultra subscribers, and in the Gemini app and AI Mode for Ultra and Pro subscribers.
$2 / $10introductory price per million input / output tokens; cached input 95% off
Speed: 288x real time on a base M3 MacBook Air (Neural Engine), against about 150x for IBM Granite Speech 5.0 TurboCTC on the same machine.
about 5%word error rate, against 2-3% for leading speech models
Setup: Claude Max and Codex Pro subscriptions, mostly Sonnet 5 plus Opus 5.5 and other models; one GitHub issue per translation unit; all agents and people in one Discord channel; Hex-Rays' ida-mcp for disassembly.
90% to 42%context-fill threshold for automatic compaction, lowered to clear stale context earlier
General health information is the most common intent in every country (28.3-48.9% of health conversations).
13-17 percentage pointsrise in the high-risk share when the scorer knows what the user revealed earlier
A 1-SD higher model score makes publication 44% more likely, or 2.9 points from a 6.5% base rate, controlling for school ranking, publication record and editorial experience.
44%higher likelihood of top-three finance journal publication per 1-SD rise in the model's score, net of author reputation
56% of first instructions in 100 SWE-chat trajectories omit requirements that later emerge; 52% of useful follow-up requirements were inferable from the project.
56%first instructions in 100 SWE-chat trajectories that omit requirements later needed
Small task (tile open windows, score out of 19): 18.3 with comments, 19.0 without; tokens 2.18M vs 1.71M; cost $1.42 vs $1.04.
7.01M to 10.64Mtoken spread across the three no-comment runs of the bigger task
The best method reached a QWK of 0.616 at Cambridge, 0.420 at Nottingham and 0.345 at Manchester Met.
761undergraduate psychology assignments from Cambridge, Manchester Metropolitan and Nottingham
52 per cent of euro-area workers use AI at work, up from 26 per cent in 2024.
52 per centeuro-area workers reporting AI use at work, up from 26 per cent in 2024 (ECB household survey)
Models are ranked by mean net units added on tasks they solved, lowest first; a second column includes failed attempts.
0.33-0.60Spearman correlation between code added and tasks solved, on boards with 26+ models
Distinction between AI-literate people, who speed up their existing workflow, and AI-native people, who redesign the workflow from the goal, including building a tool instead of buying one.
20+evaluator agents running in parallel to score one candidate's task
Faros AI 2025 ("The AI Productivity Paradox", 1,255 teams): heavy AI users closed 21% more tasks and merged 98% more PRs, while review time rose 91%, PR size 154% and bugs per developer 9%; no significant company-level effect.
+300.6%rise in time spent in QA, the largest flow-metric deterioration in Faros AI's "The Speed Trap" (September 2026)
From September, bots opened 80-90% of answered threads in the dev-request channel; every post between 21:00 and 08:00 in September came from a bot.
80-90%share of answered threads in the development-request channel opened by bots in September
Gemma4-26B-A4B ran sharded across 15 TV boxes (4GB RAM each) at 0.39 tokens per second.
8/8 nodes in 4 out of 5 trialsGLM-5.2 end-to-end replication across an emulated Android fleet with an unauthenticated port
Last 30 days: $77 in LLM spend across all agents; the secretary alone cost $51, of which $39 went on Opus answers.
88%share of the secretary's answers produced by DeepSeek, at $0.002-0.007 per answer
Consumers paid $0.37 on average for a human-directed game over its basic version, and $0.47 for AI-designed games; a 10-cent gap, 0.31 SD.
$0.47 vs $0.37average willingness to pay for AI-designed vs human-designed games (gap 0.31 SD, p < 0.001)
Same code, prompt and tools for each model; two runs each, up to 60 minutes and 30 model calls per run, with no carry-over between runs.
16candidate vulnerabilities reported across the four models
The classic FFT needs fewer than 5 n log2 n operations; the new result is asymptotically a tiny bit smaller in the log term.
10^100estimated hidden constant C in OpenAI's FFT algorithm when N = 10^12
Full score of 100 RHAE on the public set, with 7,518 actions versus 17,135 for humans (about 56% fewer) and 8,347 for the previous best entry, Sergey Rodionov's baseline1 at 99.
7518 vs 17135actions Memento 3 needed to clear all 25 ARC-AGI-3 public games versus the human total
Two MASS cycles on Qwen3.6-27B raise score per output token 1.2-1.6x on MLR-Bench, ScienceAgentBench, DSBench and AstaBench.
1.2-1.6xscore per output token after two MASS cycles versus the base Qwen3.6-27B on four research benchmarks
Two models, positioned for personal and professional agent tasks in law, finance, health and coding.
Tree docs beat skills on answer accuracy and on calling the docs or skills a question implied, in both Claude and Codex.
30%extra tokens Claude spent with tree docs compared with skills, per the tester (not seen in Codex)
Fine-tuning shifted stated beliefs about users and broader beliefs about how safe and good the world is.
1,625real-world usage scenarios used to generate all four matched training corpora
Summarising reasoning already hurt distillers: a model trained on full Opus 4.8 transcripts scored 80.4% on SWE-bench Verified, against 71.8% on summaries.
around 90 to 99%claimed cut in data harvested by distillers
CAPE-Bench: 24 task templates in eight safety families (data exports, access grants, deploy keys, backups, cost actions and others), each in two worlds that differ in one fact deciding authorisation.
70% vs 8%share of informed violations followed by a false report when the agent asked permission after acting, vs otherwise (1,081 episodes)
SWE-Maintain: 100 three-event maintenance episodes from 50 real Python repositories, with hidden tests certified to hold across each sequence.
5.6 percentage pointsaverage drop in later-task resolution when an agent builds on its own earlier changes
18 configurations: opencode and openhands scaffolds with seven models (GPT-6-Astra, GLM-5.3, Qwen3.8-Max, Grok-4.6, DeepSeek-V4-Pro, Kimi-K3, Gemini-3.1-Pro), plus four native pairs (codex, deepseek-harness, kimi, gemini).
0.5510best mean delivery score across 18 agent configurations (codex + GPT-6-Astra), vs 1.0 for the engineer reference
On 132 real, de-identified support cases, guidance was rated procedurally equivalent to ground truth on 99.2% and evidence-sufficient on 83.3%.
99.2%cases where the skill's guidance was judged procedurally equivalent to human-approved ground truth
Zero-shot small models failed: five off-the-shelf SLMs missed the quality bar, and base Qwen3-8B produced malformed tool calls on 23% of cases.
0.913agreement of a LoRA-tuned Qwen3-8B with recorded production decisions
Experts more often decomposed the problem, investigated assumptions and steered the agent; novices more often copy-pasted, delegated decisions and reviewed only on the surface.
20professionals (designers, PMs, engineers, data scientists) observed on the same task with the same agent
Survey: n = 263, 68 items.
PSS 0.813best Psychometric Similarity Score (MiniMax M3) for a synthetic sample
Early access beat late access by 0.36 points (Hedges' g = 0.62, p = 0.018) and beat no AI by 0.49 points (g = 0.83).
0.62 standard deviationsdrop in expert-rated creativity when the same 15 minutes of AI access moved from search to selection
Median tool observations are 118-286 tokens.
7.1xaverage overstatement of loop cost under flat input pricing, assuming near-ideal caching
Setup: Claude Max and Codex Pro subscriptions, default CLIs, Hex-Rays ida-mcp for disassembly, one GitHub issue per source file, agents coordinating in a Discord channel with CI failures posted by webhook.
83%functions in the reconstructed source that match the original binary byte for byte
12 models, 180 attack instances each, built from 15 APEX-Agents tasks in investment banking, law and management consulting.
8% and 20%overall attack success rates for Opus 5 and Sonnet 5, the lowest of the 12 models
Agents mostly write application code or maintenance chores; pure maintenance is 17.70% to 25.41% of agent PRs.
63.7%lower odds of human review for agent-authored PRs vs human PRs, even among contributors without write access
Some reverted features had serious architectural defects; many others had a long tail of minor edge-case bugs that would previously have been fixed gradually over about five years.
about 8 to 10significant features reverted from the PostgreSQL 19 branch after beta began
On open-source PRs, claude-opus-5-xhigh catches 79.6% of bugs humans found before merge and 64.2% of bugs found after merge, within top-5 comments.
71.9%bugs caught within the top-5 comments by the best review agent (claude-opus-5 with copilot-cli)
30% of all businesses use at least one AI technology, stable on June 2026 and up 20 points since September 2023 (9%).
30%UK businesses using at least one AI technology in September 2026, up from 9% in September 2023
Of 7,959 sites: 7,271 sent pointers (mostly x-llms-txt to llms.txt files or MCP servers), 511 policy statements, 139 detection headers, 30 injection attempts, 8 usage guidance.
7,959sites that sent at least one header aimed at AI
Russia: "phantom tenders" run to the end then cancelled for lack of budget; clients expect much lower prices because "AI makes development cheaper" and question hour-based estimates.
~5%profitability of Russian custom development in 2025, per the guest's reading of public filings, against 15-20% seen as healthy
Study 1 (explicit gender, 6,000 queries): men got higher salary recommendations in all six occupation-by-age cells; average gap $2,408 (95% CI 2,097-2,720).
$5,820higher pooled salary target recommended for male-named vs female-named candidates (Study 2)
With skills fixed, the return to AI falls with task skill and rises with each managerial ability.
Underlying GDP growth about 3%; non-farm productivity likely above 2.5% a year in 2024-2026, more than double the pre-pandemic decade.
36 percent ... Now it's 27 percentbachelor's over associate degree earnings premium at ages 22-24, then vs now
Baseline management score (2020) significantly predicts AI use in 2023 and plans to adopt in 2024, after controls for size, industry, age, region, productivity, ownership and intangible investment.
roughly 10 percentage pointshigher AI adoption in the top management-score decile vs the bottom decile
Before the mandate only 14.8% of staff had used the tool; 31.2% used it for the first time in the first mandate month and 9.1% more the next.
31 percentshare of queries that were repeated or off-task under the 200-queries-a-month mandate
Starting codex launches codex app-server --managed-daemon with no terminal, reparented to PID 1, alive until stopped or reboot.
0.157.0Codex CLI version that began starting a persistent background daemon by default
AI mentions in full-time job ads spiked in 2024Q3, almost entirely the phrase "artificial intelligence" in tech roles; 95% of that spike came from employers changing wording, and it faded fast.
3.1 percentage pointrise in the share of job ads mentioning AI per standard deviation of LLM exposure
Generative AI raises the whole quality curve: at any minute, AI-assisted work scores higher than unaided work.
219professional and student artists in the experiment (168 treatment, 51 control)
The unconstrained model (gradient boosting, protected attributes held out of inputs) is "not distribution-preserving": a non-trivial share of occupations sit above the 45-degree line, meaning amplified concentration, and amplification exceeds attenuation in size.
60intersectional groups formed from gender, nationality and age