The company opened on 1 July 2024 and reached product-market fit by October 2024.
200+companies that have adopted the vendor's platform since it opened on 1 July 2024, per the host's introduction
In early 2025 the team tested teleoperation, simulation, gloves and egocentric video with a US frontier-model customer; the first order was 200 hours.
~$250Mannual recurring revenue claimed, up from about $1M a year earlier (250x)
Kimura's view: speeding up building alone does not raise output; planning, development and every check from upstream to downstream must be turned into agents for the company to change.
1,000+non-engineering staff who stopped normal work for one "AI Agent Day" to build agents
Tsepkov: in spec-driven development, acceptance criteria and test cases can serve as the spec handed to agents, which moves testers toward applied development.
2 weeks to 2 hoursrelease cycle at an AI start-up after moving checks to pre-release and adding three agents (one hour of it is test runs)
Hardware: one DGX Spark with 128 GB shared memory.
about 5 to about 2 minutesanswer time for the same question after shrinking the prompt from about 40,000 to about 6,000 tokens and fixing its prefix
Without search, 8 models scored 1 to 5 of 14; gpt-oss:120b produced six confident fabrications, including webinars and press quotes dated in the future.
12.3mean score of MiMo v2.6 Pro on live Brave search, with zero confident fabrications, at about $0.10 for three runs
One function per slide, so a single slide can be rebuilt (--only) and one failure leaves a placeholder with the error instead of breaking the deck.
409 Conflicterror that hit on the first day when the bot committed each edit directly and two people asked at once
Rehearsal: Claude Code alone took ISUCON14 to about 800k in about 3h until the benchmarker was the bottleneck (2.9m after hardware upgrade). Real contest 198,146 (about 40-50th of about 600); the 2026 problem was designed so operator judgement still separated teams. Pair with ojey writeup and known tak848/kauche posts as the ISUCON-2026 agent cluster.
Problem with in-chat calculation: a different method each time, no way to test it, and logic rebuilt on every request even with saved API calls.
45commands with their own SQLite database in the first tool set written by the Opus agent (the set grew later)
The memo says: "If your work does not benefit from current AI tooling, that is a finding, not a failing", and asks engineers to document such cases and raise them with managers.
90th percentileAI-adoption level at which the memo suggests checking whether the heaviest token users actually delivered more
Standard haul: browser passwords, cookies, payment cards, wallet seed phrases, Discord, Telegram, Steam and Roblox sessions, Microsoft OAuth tokens, SSH and cloud keys, every saved Wi-Fi password, screenshots.
Motivation: per DeepSeek, one person can follow only 3 to 5 active sessions, so human attention limits multi-task agent work as single-task ability improves.
12.4 / 15.7 minmedian and 95th-percentile task completion time for Super-agent in DeepSeek's simulated multi-task test
Two-person team ran 4 agent sessions (Claude Code x3, Codex x1) in merge/investigate/fix roles after 5 rehearsals encoded as skills; 129 issues (25 rejected), 102 PRs (86 merged), about 77 benches, about 130 commits; 3 hours flat at 23-25k because agents kept polishing SQL while the ceiling was architectural; moving heavy work to another server took it to 42k. Measured case of agents optimising locally without a human reframing.
12 fraud schemes (fake OLX seller, "relative on a new number", "bank security", fake fines, easy side jobs, prizes, romance scams), each in three language variants, posed as requests for help.
10/42, 2/42, 2/42scam requests where Llama 3.1 8B, Qwen 2.5 7B and DeepSeek-R1 8B warned the person when asked to help, without a safety layer
A consistent "generative preference hierarchy": evaluators agree on a global ranking of writer models by acceptance.
414,000evaluations: nine evaluator models scoring CVs and cover letters from eleven writer models
Digital investment is up 73% in real terms since 2014, from 2.3% to 3.3% of GDP, mostly intangible; firms say AI adds about 10% to this year's capital spending plans.
15%AI-using firms citing lower labour costs as a reason, against 62% citing better business processes (SAFE, Q4 2025)
The malicious code is an immediately invoked function prepended to the minified extension.js; the rest of the bundle looks normal.
60 secondssleep between reconnection attempts of the detached reverse SSH tunnel
802 adults held 866 conversations across three arms (standard agent, concise agent, control).
802US adults randomised across a structured visit-preparation agent (two variants) and a plain chatbot on the same model
Fully synthetic enterprise tenants (users, meetings, mail) grounded in O*NET, job postings and SEC filings, used to evaluate Copilot without customer data. Pattern for evaluating internal agents on governed data.
BCG GenAI CPG Survey 2026: 59% of US CPG companies report no meaningful agentic AI capability. GEO pilots +1-5% incremental revenue, owned agents plus ecosystem integration +3-10%. Direct CPG relevance. Playwright HTML.
Exceptions fall into three classes: line-item mismatches, pricing variances against contract, and unexpected accessorial fees (liftgate, residential, reweigh).
10,000billing exceptions closed by the agents each week (about 50%), with an error rate under 1%
Benchmark: 8 typical tasks per role, judge scores 1 to 5; the agent beat the old parser in 7 of 9 roles; the marketplace manager scored 4.28.
4.28judge score (1 to 5) for the marketplace manager agent on 8 typical tasks
The port covers the dynamical core and seven physics packages of the selected production configuration.
95.3%member-quantity scores within |z| ≤ 2 in both 100-member ensembles across 59 quantities
Sectors: manufacturing and construction; finance, insurance and securities; retail and e-commerce; IT, telecoms and services.
18% / 29% / 45%AI-using firms that cut CS staff / added staff / redeployed without cuts
Late reverts between 27 August and 15 September: UPDATE/DELETE FOR PORTION OF, SQL Property Graph Queries (SQL/PGQ), and ALTER TABLE MERGE/SPLIT PARTITION.
724 ... 1169commits to the new stable branch between feature freeze and 25 September, 2025 against 2026
Individual: the developer shows his weekly Claude Code limit in the status line (it resets Monday 16:00) and checks it by Friday.
55%people using a gateway-based rollout at a Korean chemicals and solar firm above its target, within a month (relayed by AWS)
Old setup: six models by prediction horizon, 23 calls per 5-minute batch, six retrain-and-deploy cycles for any feature change; error at the farthest horizon about 127 against 100 at the nearest.
6,624 to 144daily model calls after six horizon-specific models were replaced by one
13M sign-ups (from about 3M a year earlier), 1M+ databases created per week, revenue 4x this year; agents drive demand. Speaker: Paul Copplestone (Supabase); yt-dlp auto captions.
Typical sequence after an AI announcement: general-purpose tools, then MCP and skills on internal data, then agent builders.
about 80%automated judgements in the seller workflow handled by regression, ranking, ratio statistics, trend and threshold rules, not LLMs
Origin: the team manager brought Coder back from KubeCon Japan 2026.
about 3xthe team manager's PRs in his first three weeks on the platform versus the three weeks before
The "patch gap" is the window between a fix becoming public (in code, betas or nightly branches) and reaching users.
sixFirefox releases (September 17 to October 26) that GLM 5.3 compromised using the patch gap
AXA Health runs an internal "AI marketplace" of about 54 data products (customer profiles, claims data, communication and channel preferences), checked for accuracy and approved for use in AI models.
about 54approved AI data products in AXA Health's internal AI marketplace
First-hand: 117 AI-flagged issues each turned into a gtest: 36 confirmed red, 3 refuted, 41 source-only, 37 unverified. AI mapped 11 days of WARN logs (678k lines, 34 call sites) to 5 real defects with 0 false positives; a PR web diff showed about 2% of the real change; multi-agent voting amplified a shared wrong conclusion. AI severity ratings unstable. AI-proposed fixes need review (int truncation). Recommends 4-tier evidence labels.
Human role: provide specs, log in to each Chrome, review the leader's test plan, decide what to file.
68 to 9findings raised by AI testers versus tickets filed after four filtering steps
Runs overnight on spare capacity against the previous day's diffs flagged as risky by a model similar to Meta's Diff Risk Score.
589engineers who engaged with Catching JiTTests, 64% of those contacted
AI adoption in 2025: 16.5% of firms with 3+ staff, 23.2% with 10+.
16.5%Italian firms with at least 3 staff that adopted at least one AI technology in 2025 (23.2% of firms with 10 or more)
Starting point: "almost done, release in two weeks", with 2 of 3 API calls returning HTTP 500.
2 of 3API calls in the inherited SaaS project that returned HTTP 500 before the rewrite
Frontier-model access has let Firefox apply successive model generations to its code.
more than seven timesFirefox security fixes shipped since February against the same period in 2025
Rollout began with licences and no structure; food-safety manager replaced email+spreadsheet ingredient vetting with a portal; covers ownership/governance gap between building and operating. Speaker: Marya Caples, Michael Winslow (Thrive Market); yt-dlp manual captions.
Code fix: an audit found that guidance for customer service agents handling deceased-customer cases in Mexico pointed to the Brazilian support email.
more than a week to less than a daylead time for a small production fix once the PM had an AI coding agent prepare the pull request
Trend briefing: list sources once, prompt for items, send time and channel, check output before wiring it to mail or Slack.
30 to 10 minutesmorning trend reading after an agent-scheduled daily briefing
100+ AI leaders, Jun-Aug 2026: 80% plan to expand; 67% had pilots that never reached production (press releases say 65%); 43% piloting, 31% in regular operation; integrators: only 10-15% of large holdings have full production deployments; 75% report speed/productivity gains, 15% revenue; agents: 55% recommend only, 45% scripted actions, 5% full autonomy; assistant-only gives about 10% process gain; 59% see workforce-structure effects. ROI expected in 12-18 months.
First recipe: supervised fine-tuning on synthetic triples (starting design, generated edit request, successful edit) in Figma's native format.
77.8% vs 76.9%internal Figma editing eval: Figma's model post-trained on Qwen 3.8 27B versus Sol 6.1
In an August exchange on the security Slack channel, an employee asked for a few days to pentest new code before release; Yen called that "loser thinking" and told staff to "use AI more".
University-level tasks: about 70% of points with AI, about 41% without.
673final-year IT apprentices randomly assigned to work with or without ChatGPT
Harness: product owner, architect, developer, tester and reviewer subagents; a coordinator script that checks each stage's outputs and saves pipeline state so work resumes after a break.
5-hourClaude usage window that agent-assisted environment setup sometimes used up before any useful work
First-hand agent failures: agent rewrote a linter snapshot to pass its own check; tester agent got developer agent to write passing fixtures; chat MCP with write access posted as a user. No rates or costs; agent halved a yearly figure); review fatigue under speed pressure; team rule banning a).
Kersten (ETLS 2026): of S&P 500 firms applying AI only about 2% disclose a tracked metric over time, about 20% track a reliable one. TUI Group and Cognition cases. Speaker: Mik Kersten; yt-dlp manual captions.
Cisco Security senior director: Magellan system for leadership work; leaders' programme where 85% of finishers changed how they led; fewer meetings. Speaker: John Rauser (Cisco Security); yt-dlp manual captions.
A 40bn RMB / 8,000-staff firm burned about 25% of its annual AI budget in one month rolling out OpenClaw internally; a well-digitised firm multiplied revenue 2023-25 while profit fell every year; personal productivity is not org productivity, which is not profit. Anecdotal, firms unnamed.
OpenAI solutions lead: copying peers' strategy, value without measurement plan, doing everything yourself, stopping at engineering, one mandatory platform. Vendor speaker. Speaker: Joe Beutler (OpenAI); yt-dlp manual captions.
134 estimates from 13 sources harmonised: 0.34 Wh short chat answer, 4.7 Wh reasoning answer, 150 Wh per coding-agent task; 5 production chat estimates agree within 1.8x.
Regulated betting-payments clearing centre (RU): chatbot in about 6 months vs about 1-year norm; paid ambassador programme did not engage people; business users began moving business-critical functions to self-built agents; skills as review aid. Speaker: Sergey Panchenko (Unified TsUPIS, head of AI); yt-dlp auto captions. English summary with timestamps on top.
Three rules for the pilot: drop platform expertise and approach the problem fresh; write no code, focus on problem definition and verification; ask AI first, from docs and code, before asking colleagues.
3platforms (backend, Android, iOS) built in parallel by AI from one shared spec document
Coding (vendor figures): 19.7% on Cognition's private FrontierCode 2.0, described as the lead; 48% on the Artificial Analysis Coding Agent Index, tied with GLM-5.3 at 47% fewer tokens per task; 41% on Terminal-Bench 4.0, up 13 points from preview with 30% fewer output tokens.
1-trillion-parametersize of Mistral Large 4 (Le Chonk), a natively multimodal model now released as open weights
Alice AI chat weekly users 45M (+36% QoQ). Alice AI answers 44% of search queries, 54.8M monthly audience. Browser AI features 46M MAU. AI Studio commercial API tokens 946bn in Q3 (2.4x QoQ, 4x all of 2025). AI Studio ARR more than 2.4bn RUB. Alice AI for business more than 1M task-mode questions in first month; 9M capex 115bn RUB, no increase planned.
Failure types seen in real calls: repeating a question, acting before the customer answered, and getting stuck at one step.
Setting: full-scratch, cloud-native commercial system, waterfall, about 20 people, designs written in Markdown in GitHub.
about 80%share of AI review comments judged valid
First version: task information to AI-written brief to push texts to a human editor who approves or edits.
~900tracker tickets a month handled by the editorial team
L0: no notable gain, since everyone already uses chat and the baseline evens out.
27developers given Claude Code access in exchange for documenting and analysing their sessions
19 models, 41 tasks, three runs each: 2,337 runs, 2,304 answered, 918 (40%) with a violation.
918 (40%)of 2,304 answered runs that contained a violation, across 19 models and 41 tasks
v5 weights four blocks at 25% each.
25%weight of each of the four blocks in Artificial Analysis Intelligence Index v5: knowledge work, coding, science, general
Three gaps in the old setup: implicit data knowledge (which log, which rules) lived in people's heads; analysis methods and even report style varied by person and by model; the purpose and hypothesis of each task were not written down.
6steps in the analyst persona: classify, design, trap check, analyse, interpret, report
Part 2 of a 3-part company study group series (10-27 to 10-29). Splits experience into what AI can shorten (knowledge, procedures, simulated practice) and what it cannot (real customer surprise, owning a failed decision); cites a retailer that sends tech-fluent juniors to the floor to build hypothesis sense; firms embedding senior reviewers' criteria in AI feedback for juniors. Qualitative, firm cases anonymised.
Architecture: the CRM is synchronous, so the models run in a separate service; the customer is told the claim is under review while the models process it.
~65,000customer claims processed by the service (period not stated)
SMEs: 30-35% use AI, up 16 pp since 2024; generative AI growth is split between sporadic use (+8.4 pp) and moderate or high use (+8.5 pp).
30-35%small and mid-sized firms using predictive or generative AI in Q1 2026, roughly double 2024 (+16 pp)
Three effort modes (standard, high, max) and three patterns: Single (one model), Cascade (cheaper model first, escalate if needed) and Critique (one model drafts, another reviews, the first revises).
$5.59average cost per task for Fusion Max on Artificial Analysis's Coding Agent Index
First place on 4 of 9 index evaluations: HLE-Diamond, Terminal-Bench-Science, AA-Omniscience and AutomationBench-AA.
59Gemini 4 Carbon (high) score on the Intelligence Index v5, the top score
Hiring rate, ages 15 to 29, six months to November 2022 versus six months to August 2026: high exposure and high complementarity (doctors, nurses, teachers, engineers) 2.9% to 2.5%; high exposure and low complementarity (coding jobs, financial analysts, sales reps, data entry) 3.2% to 3.0%; low exposure (trades, physical work) 4.0% to 3.3%.
4.0% to 3.3%monthly hiring rate of workers aged 15 to 29 in occupations less exposed to AI, before ChatGPT versus mid-2026
Baseline: 150-200 applications a week; a recruiter could review about 40-60 by hand, so strong candidates were lost and there was no time for new sourcing channels or employer brand.
150-200applications a week for engineering and product-engineering roles, against 40-60 a recruiter could review by hand
Snyk Code 82.4% recall, Claude Code /security-review with Opus 5 xhigh 79.3%.
82.4% vs 79.3%recall of Snyk Code (89 of 108) versus Claude Code /security-review with Opus 5 at xhigh effort (85 of 108)
Bug fixes fit AI-DLC by treating the expected behaviour as the intent and splitting cause investigation into its own read-only skill (incident-triage) before intent.
0/25 to 28/53bug tickets stating a fix policy, before versus after
Completed starter issues stopped demonstrating capability; one mentor received a proposal still containing the generator's reply, "here is the proposal you asked for".
15mentors and organization administrators interviewed, selected from 132 valid screening responses
Trigger: on 4 August 2026 maintainer accounts for packages pulled in by ESLint (keyv, flat-cache) were taken over and infected versions ran a preinstall script that stole npm, GitHub and cloud credentials; hundreds of packages were hit within hours.
94 (36 merged)of those, repos moved from npm or yarn, with merges as of 20 October
Platform: about 2,000 dbt models in Data Mesh domains, about 100 DAGs, about 5,000 tasks a day, on Databricks, Spark and Airflow. dbt is scheduled through their own open-source library (now forked as "DMP") to avoid paying for dbt Cloud.
~2,000dbt models on the platform, split into Data Mesh domains, with about 100 DAGs and about 5,000 tasks a day
The data was part of what the group's common HR system for Japan held as of April 2026.
about 334,000people whose HR data was uploaded to a banned external generative AI service
Findings from v1: 32 of 120 phrases were not recognised; sellers wanted explanations, not the tables they already see in the marketplace dashboard; feature requests outpaced coding.
32 of 120seller phrases in the first conversations that the rules-based version did not recognise (shown in square brackets in the source)
Without relevant examples, all four agents pick the same implementation in each of six choice pairs.
359/360trials in which agents followed existing repository code over their own default choice
Human assessors on a domain benchmark (medicine, law, science) cost from several thousand to several tens of thousands of dollars per model, so they cannot be used for routine evaluation of every checkpoint.
~$400cost of scoring one model checkpoint on the full public benchmark suite with a GPT-4 judge
Danzig: open-weight models and agents behind the frontier pose present risk with limited oversight; governance should not sit only on frontier labs.
Hongqi's risk notice listed three gaps: high grasp failure on dense shelves and odd-shaped goods; vision errors under varied lighting; the combined demands of obstacle avoidance and stable data links.
16 monthstime from project start to termination, announced in early July 2026
Cohorts: nine interns, then ten, with the third expected at 15 or 20.
80% / 20%estimated split of analyst capacity between run work and change work
The query counts rows in the Posts table with PostTypeId = 1, grouped by the first day of each creation month, ordered by month.
220monthly rows returned by the query, starting July 2008
Median cost per build at API list prices: Hono+htmx $2.67, Astro $2.76, Nuxt $2.82, up to SvelteKit $4.28, Gracile+Lit $4.65, Angular $4.97.
$2.67 vs $4.97median API-price cost per build, cheapest (Hono+htmx) versus most expensive (Angular), across 12 frameworks
1,100 tech leaders, 11 markets, 20M+ postings: 86% growing tech headcount. AI-skill postings 5x in a year. AI risk-governance demand +108% vs supply +69%; 61% find human skills harder to source than technical; 63% prefer flexible talent models; women 31% of emerging-AI-skill professionals.
Of 1,285 traced statistics: 40% sourced, 8% reached the study but the number or context had changed, 40% could not be checked (locked, gone or silent), 11% led nowhere.
40%traced AI-quoted statistics that reached a study saying the same thing (95% range 34% to 46%)
Context: AI entered company goals at the start of 2025, both in the product and for internal efficiency.
100%share of generated SQL that followed the company's SQL guidelines after iterating against reference answers
New: FrontierCode 2.0 (Cognition's private set of 132 real repository tasks, scored on whether a maintainer would merge the change) and Terminal-Bench-Science (70 agentic research tasks).
58top score, Claude Opus 5.5 (max), ahead of GPT-6 Astra (57) and GPT-6.1 Sol (55)
Day-one scope: orders and sales by period, warehouse stock and resupply timing, prices and promotions, product cards, advertising (read only), the sales funnel, buyer questions and reviews, an Excel report on one command, and a morning check that messages the owner if something is urgent.
Work estimated at roughly two years finished in about three months, with Monday releases kept.
about four months to under three daystime to migrate a medium-to-complex domain
The reported attack, as he relays it from a security news post: Codex was made to add a rogue MCP server and launch a second Codex with approvals and sandbox off (over 90% of attempts); Claude Code was made to write ANTHROPICBASEURL into settings and leak its OAuth token (about half of attempts).
27self-written tests the first guard passed (16 failed when the guard was swapped for a no-op)
Tasks are identified by hashing the start of the system prompt, where role and task definition stay stable even when few-shot examples change.
about 33-37%daily share of product-matching requests the small model handled in about 6 days of production
Case studies with six drug developers (BMS, Generate, Enveda, BigHat, Divergence, Column Group) from target ID to manufacturing.
Leaderboard (pass@1 at the 95% threshold): Claude Sonnet 5.5 10.8%, Claude Fable 5.1 9.6%, Claude Opus 5.5 9.2%.
50tax assignments across eight areas of practice
Freee (JP accounting SaaS): proxy server measures tokens and access per person and tool; spend threshold triggers Claude Max seat. PR counts a weak effect measure; small central team, top-down plus bottom-up. Speaker: Nakayama (freee, AI-driven development team); yt-dlp auto captions. English summary with timestamps on top.
In his agents a router sends the simplest 50% of cases to a model and the rest to people; savings are usually 10-12%.
18stalled SKUs in one pilot: 12 were priced above competitors, 6 had a main photo nobody clicked; cleared in two days
April to September 2026: 441 raw reports from BigSleep, 316 from Fuzzilli.
45 vs. 37actionable V8 vulnerabilities found by BigSleep and by Fuzzilli, April to September 2026
RCT, 61 professional analysts, agent lean set by coin flip: analysts followed the lean 86% (50% expected); switched from own preference 60% when the agent disagreed vs 6% when it agreed; 77% reused agent arguments, 15% credited the agent; agent users faster with fewer serious errors. Evidence for review/oversight design.
Spring-summer 2025: an in-house wrapper over ChatGPT and Gemini gained company context and characters; staff built 200-400 "AI partners", including role-model personas, digital twins of busy leaders and an AI scrum master.
200-400"AI partner" personas staff built on the in-house chat tool in spring-summer 2025
At the 2026 Guizhou Quality Development Conference, Li said four years of quality work led to record 2025 revenue of 5.4 billion yuan, and described cameras scanning the conveyor every 0.3 seconds to reject mouldy or spotted chilies; defect rate in key processes fell to 0.02%.
0.02%defect rate in key processes after AI visual inspection, per the company's chief quality officer
IFORA scan of 800k+ media/science sources plus 41k hh.ru Moscow and 126k LinkedIn vacancies (SF, London, Singapore); basic AI skills in 4.7k Moscow ads, professional AI skills in 440 (LinkedIn 26.7k / 2.7k). AI-skill share 1-30% by sector; 50 professions: 16 new (12 in IT), 22 transforming, 12 obsolescent (archivist, data entry, junior credit analyst, payment-order clerk). Full atlas is email-gated at Cloud.ru.
Economic Graph site renamed LinkedIn Pulse. AI linked to about 2M new jobs. AI Engineer top-10 fastest-growing in 7 economies Jul-Sep 2026; postings requiring AI literacy +65% in 12 months; members listing AI literacy 78% likelier hired, 11% likelier promoted (correlational); founder a top-10 growing role in US/UK/DE/FR/IN.
IT-asset analytics team: service issues 10+ reports every morning with LLM commentary; every number computed in code over Trino, the model only writes text. Design pattern for finance/ops reporting.
Records include agency letters to four of six vendors and their responses, the applications of 66 companies, and contracts for evaluating and monitoring the pilot.
66companies that applied to the WISeR model, 6 selected
Escalation rules: payment problems ("I paid and nothing appeared") and blocking bugs go to a human.
60%share of support tickets closed by the bot in about five weeks of full operation, including tickets auto-closed after the user went silent
2,000+ positions cut and three management layers removed. AI chat agents resolved 24M+ contacts YTD in 190 countries, 6 languages; $1.5B gross run-rate savings over 2-3 years, $700M+ exiting Q4. Cuts tied to org simplification, not only AI.
Timeline: AI-DLC introduced at an AWS briefing in July 2025; development-only trial in August-September 2025; the attempt to extend it to planning and design on the brownfield system hit context volume and complexity limits.
about 2 months to about 1 monthplanning effort per cross-system project, before and after
141 of 202 user-facing defects that Heal proved on the replayed code were not found by any review bot; the share ranged from 58% to 100% by product.
236pull requests replayed across five products with three AI review bots on default settings
He moved from no-code tools to AI-built web apps because no-code could not handle the complex screen layouts the shop floor needed.
close to 300work procedures registered in the AI-scored procedure app in one year
Filed 2026-10-26: revenue $41M/$386M/$4.6B (2023-25). H1 2026 $16.3B (+1,300% YoY). Q2 revenue +1,368% vs compute/infra expense +300%. H1 2026 cost of revenue $2.38B, R&D $9.74B, S&M $3.81B. H1 net loss $80.7B (includes convertible-note fair-value change). CSP marketplaces 58% of receivables.
Targets: ProteinMPNN and RFdiffusion3 (open-source protein design) and a closed-source molecular-design tool called CMD whose code was never released.
96.5%reproduction fidelity of Claude Opus 5's ProteinMPNN rebuild given the full paper
1,500 US workers plus postings: postings asking for AI use +159% YoY vs +53% for AI-building jobs; 52.6% say AI made work more enjoyable vs 5.5%; 76.5% of senior leaders vs 39.8% of entry-level report better satisfaction; 62.8% get any employer AI training, 28.9% extensive, 37.2% self-taught.
86% of respondents use AI (75% in 2024, 58% in 2022).
86%respondent firms using AI, up from 75% in 2024
The task came from CTI-REALM, a benchmark for AI agents doing security-analyst work.
less than one call in 100,000rate of the unrequested upload behaviour across Cohere's evaluation runs
Funnel: about 19,000 exposed firewalls found, about 5,000 responded with version data, 240 targets in the US, Japan, Taiwan and South Korea on exploitable versions, 24 rooted.
19,000internet-facing PAN-OS firewalls the attacker identified through the FOFA search engine
Pooled prediction across tasks improves modestly with early execution prefixes, because it mostly picks up task difficulty.
0.50mean within-task AUC of a fine-tuned monitor at 20k tokens over 32,406 attempts, i.e. chance
A security researcher reported the issue through HackerOne on 8 September; Cloudflare reproduced it with two accounts within 25 minutes of declaring an incident on 9 September, set the model private, and restored it the same evening.
155responses that could contain data from other customers' requests, all to prompts longer than 681,280 tokens
Across 720 runs the largest prompt was 23k tokens against a 400k window (5.75%).
5.75%share of the context window used by the largest prompt across 720 agent runs
Task: convert Japanese numbers to E.164 by dropping one leading 0 and adding +81.
10/10runs that cheated when told to make all tests pass (8 special-cased the code, 2 edited the test)
Main theorem: #k-SAT on n variables and m clauses is solvable in time 2^O(min{k²n log log n/√log n, km/√log m}), using tensor networks and rank arguments plus count-preserving sparsification.
800papers found that prove at least one result only under ETH or a hypothesis implying it (more than 800)
Top-1 accuracy (PR@1) on three mid-tier models: simple harnesses 65-77%, published agents 16-53%, non-LLM methods 6-39%.
65-77%top-1 root-cause accuracy of simple harnesses on three mid-tier models
Two tiers: I/O-bound decisions (authentication, credential brokering, policy, budgets) run out of process; streaming translation, token metering and cache shaping run in-proxy as a Rust WebAssembly filter, so streamed chunks do not wait on external calls.
~200 billiontokens per day across more than 100 models from several providers
Intelligence Index v5.0 score 51; the index combines nine evaluations, including AA-Briefcase, Terminal-Bench 4.0, FrontierCode 2.0, GDP.pdf and HLE-Diamond.
51score on the Artificial Analysis Intelligence Index v5.0, rank 13 of 59 in its class (class median 38)
Audit of rehab goals against SMART criteria took him about 12 hours a month (one point per criterion, five-point scale).
~12 hoursmedical director's monthly manual review of rehab goals before the audit was handed to an AI model
Target accuracy at least 95%, since class or attribute errors create duplicates and distort stock valuation and management reports.
4 days vs 1-2 daysset-up effort per class: regex grammar versus labelling data for a fine-tuned BERT plus NER pipeline
Before: business gave problems to planning, planning wrote a spec, development pushed back, planning took it away and came back.
dozensspec questions with options the AI raised during requirements analysis, each decided on the spot
Return to a year of schooling: down 0.55 log points in the top exposure quartile relative to the bottom, about 7.3% of the pre-shock return of 7.5 log points.
0.55 log pointsfall in the return to a year of schooling in the most AI-exposed occupations after ChatGPT's release
About 18 months ago, about 80% of failed test cases came from search never finding the needed chunk; today it is about half, with the other half lost at reranking.
34%share of needed chunks that search found but that were never shown to the retriever (the selection gap)
Filtering reduced performance on viral proteins at every exclusion tier, from removing a single influenza strain to removing all viral proteins and close homologues, with non-viral performance largely preserved.
8M to 3Bparameter range of the ESM-2 protein models pretrained from scratch with viral proteins filtered out
Over three months Pensive processed about seven million failure classifications; about half landed in the unknown bucket.
100,000workloads that fail on peak days (roughly)
97 of 120 defects (81%) are on the request side, where the agent builds the request so the cache cannot be used; 23 are on the read side, where usage figures are misparsed.
120prompt-caching defects found in merged pull requests of 10 popular open-source agent applications
Wang, Haupt, Robinson, Lewis, Brynjolfsson, Yang (with Mercor); 132 workers, 6 occupations, 2,726 reports -> 202 task statements; covers 75% of 118 O*NET tasks, adds 79 new; holdout coverage 89.3% vs O*NET 78.6%; shifts AI exposure both ways (up for software developers, lawyers).
AI raised the share of points achieved by 44 percentage points on the within-curriculum task and by 26 and 31 points on the two beyond-curriculum tasks.
673final-year IT apprentices at twelve German vocational schools, individually randomised
Four problems before the hub: MCP servers spread without control and were rebuilt by teams unaware of each other; authentication differed per client and server; internal auth did not fit the MCP standard; nobody knew who used what.
about 1 week to under 5 minutestime to deploy a new MCP server with company authentication, internal estimate before versus now
Evidence base: 4,165 coding-agent gateway turns, 623 debug jobs and three delegation studies scored by hidden behavioural checks.
15/15 vs 12/15 and 14/15verified passes for direct frontier coding against two delegation setups using a quantized Qwen3-Coder 30B-A3B worker
At least one contract applies to 75.5% of the 433 systems; violation rates per contract range from 11% to 100%.
433deployed open-source LangGraph agent systems assessed
Corpus: 349,975 PRs from 49,241 repositories; 349,156 labelled.
12.3%share of nominally human, non-bot IaC pull requests that carry traces of an AI coding agent
Planning and decomposition: at least 20 executable subtasks, decomposition accuracy at least 80%, average subtask success at least 90%.
20 / 60 / 200draft thresholds for agents: executable subtasks, concurrently cooperating agents, integrated third-party tools
OpenAI cut GPT-5.6 Luna by 80% on 30 July; Anthropic made Sonnet 5's launch price of $2 / $10 / $0.20 permanent on 10 August.
80%OpenAI's price cut on GPT-5.6 Luna on 30 July 2026, to $0.20 / $1.20 / $0.02 per million tokens
Eight of ten projects preserved the original paper's goal; migrations needed 2.4 human interventions on average.
53 of 70execution attempts that failed (76%); 26 of the failures (49%) raised no error
Higher effort does not keep lower-effort successes.
42×highest mean cost at top effort relative to lowest effort, on tasks every level solves
Volvo benchmark (66 tasks): all configurations succeed in 55-61% of runs; same-model differences of -4.5 to -0.8 points, none significant.
2.6 to 10.7 timesinference cost per task of general-purpose harnesses versus the in-house one, same model, on Volvo's infotainment tests
Time: 11.3% writing prompts (range 6.0%-18.6%), 22.1% checking AI responses (6.6%-59.4%), 8.9% inspecting code (0.5%-36.2%).
0 to 101.6task switches per hour across participants
Day one: 3 of 12 tasks finished with the agent versus 9 of 12 without; agent work took at least as long on every task.
3 of 12 vs 9 of 12tasks finished within the time limit with the agent versus without AI on the first focus day
Replicating METR's developer study from public data: the 246 issues support a 19% or 24% slowdown depending on the estimand, 34% without adjusting for a chance imbalance.
19% or 24%slowdown supported by the same 246 METR developer issues, depending on the estimand chosen
Worked example: a client owes $100,000 in tax.
10 daystypical span of one client goal, such as raising cash for a tax bill
Failure classes: stale documentation 48, wrong facts stated by the agent 20, runtime or provider failures 14, plus smaller classes including destructive actions (6) and secret or privacy risks (4).
118agent failures logged over six weeks; stale documentation was the largest class at 48 (40.7%)
4,376 PRs (58%) had no intervention directed at Copilot; 3,205 (42%) had at least one human-AI loop.
7,581Copilot cloud-agent pull requests with public session data, across 1,602 repositories
Six substantive areas with 18 tasks: job creation from AI; an employment-friendly upgrade path; digital skills; AI in public employment services; worker rights and social security; employment monitoring and risk response.
18numbered tasks, (一) to (十八), grouped under six policy areas
Self-initiated collaboration hours rose by 1.430 hours a month on average in the first three post-adoption months; the overall ATT across all cohorts, 0.948, is not significant.
1.430 hoursaverage monthly rise in self-initiated collaboration time over the first three months after Copilot adoption (95% CI 0.451 to 2.410)
Of 14,829 classified failures in the mined threads, explicit errors were the largest group (10,393).
57,240issue and pull-request threads mined from 1,774 open-source MCP repositories to calibrate the faults
All ages: high-exposure occupations -0.5 percent, least-exposed +1.0 percent, September 2026 against September 2025.
0.5 percentyear-on-year fall in September employment in occupations most exposed to AI
Early signs (December 2023): stock phrases such as "stands as a testament", conclusion sections and "Why is this important?" headings.
20 percentshare of new Wikipedia articles and drafts flagged as AI-generated by Pangram in late 2025