AI moves fast. We slow down just enough to see what matters. 60 source-based analyses through the Cortist Lens.
The AI scored about 83% across 304 cases; reference-assisted generalists scored about 36% on a smaller 56-case subset. This does not establish safe clinical use.
83%Retrospective diagnostic accuracy
948working LPE exploits (company report)OpenAI says its cyber-agent pipeline reproduced 1,724 Linux flaws and demonstrated 948 local root exploits for $1.21M. Maintainer-verified fixes remain a separate question.
21MRNA sequence–structure training pairsNature reports structRFM, a sequence-and-structure RNA model trained on 21M pairs with open weights and code. Selected structural benchmarks improve; therapeutic benefit is unproven.
$33.1Bestimated proposed investmentLambda plans a $33.1B Alabama AI campus, with power approval pending and tax, jobs and water claims still forecasts. We separate announced commitments from delivered outcomes.
82.2best teamwork scoreA 240-task benchmark of 18 AI models finds that the best agent teammate scores 82.2/100, while conflict and corrupted handoffs expose reliability gaps.
7×upper cost disparityNVIDIA researchers report up to a sevenfold late-token serving-cost gap and little loss from frequent context compaction on selected tasks. The claim needs replication.
$895Bcompute commitmentsAnthropic’s November 2 S-1 amendment adds Q3 figures and $895.4B in compute-heavy commitments. Cortist separates obligations, costs and profit.
1,000MW planned flexibilityOpenAI’s Grid Protectors program cites a 1,000 MW Georgia demand-response agreement. The capacity is contracted, not a measured delivery result.
66%lab task successBioAgent reports 66% and 58% complete-task success across two real-robot lab protocols with recovery—promising, but not autonomous science.
7×detected missed referralsIn seven Singapore clinics, MR-AI raised missed-referral detection sevenfold, but 48.8% sensitivity shows why human follow-up remains essential.
20hidden-rule chainsA new hidden-rule benchmark finds that many self-evolving agents memorize feedback rather than generalize; stronger models make a measurable difference.
27.5percentage-point maximum gapBlindToChange finds gaps up to 27.5 percentage points when agents must discover recent changes themselves. Cortist examines verification.
13/93full reproductions in top runPaperBenchX tests 93 scientific experiments across 12 areas. The top agent fully reproduced only 13; Cortist examines clean replay.
55%compute expense / H1 revenueAnthropic's SEC S-1 reports $16.286B revenue and $8.945B compute expense over six months. Cortist separates growth from accounting losses.
1stGoogle-reported data-center TPU response in orbitGoogle reports its first AI response from a data-center-class TPU in orbit. Cortist separates the Project Suncatcher prototype milestone from viable space-based AI infrastructure.
88.2%SEC-bench Pro — vulnerability findingGoogle reports 88.2% on SEC-bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench. We separate different tasks, access limits and actual autonomy evidence.
29controlled Unitree G1 jointsMetatron v1 publishes native and browser model artifacts for Unitree G1 whole-body control. Cortist reviews the 29-joint interface, weight licensing and evidence gaps.
35 sReported fight-to-dance transition on a real Unitree G1A 27 October research paper reports generated 16.8-second behaviors and a 35-second fight-to-dance sequence on a physical Unitree G1. Cortist separates motion tracking from industrial autonomy.
56 → 75%Company-reported success rate in one real-world manipulation eGoogle says Gemini 4 Argon used a simulation twin and more than 4,500 synthetic episodes to improve a real-world robot task. The results are company-reported and task-specific.
1,250 hBMW-reported cumulative operating time during its 2025 pilotBMW's 2026 retrospective describes a 2025 Figure 02 pilot that handled more than 90,000 sheet-metal parts. Cortist tests the difference between cumulative hours and independently audited shifts.
9.1 moFitted doubling time for quarterly H100-equivalent chip sEpoch AI estimates 2.5× yearly growth in compute capacity sold between early 2024 and mid-2026. These are chip-sales estimates, not proof of deployed capacity or AI progress.
6 / 6Historical Firefox releases compromised in controlled tesEpoch AI reports that GLM-5.3 built successful attacks in controlled tests of six older Firefox releases. Mozilla describes mitigation; the result is not evidence of attacks on users.
30 minTime to warning in a replay of a prior recorded incidentOpenAI reports a wider system for monitoring frontier-agent misconduct, a simulated earlier alert and a live training alert. Cortist distinguishes detection, response and prevention.
72,191Turing-like judgments in a peer-reviewed research studyA peer-reviewed Nature Human Behaviour study tests how well people identify AI-generated responses. Cortist separates imitation from intelligence and real-world deception.
0 / 135AI-generated follow-up papers meeting every domain expe…A national-laboratory research benchmark found that 114 of 135 AI-generated scientific follow-ups failed an integrity check. Cortist explains the limits of this preprint.
3 claimsMathematics, complexity theory and quantum field theoryOpenAI released AI-produced mathematical and complexity-theory proofs and described a physics collaboration. Cortist checks what is public, formalized and not yet independently settled.
39%Shutdown-script changes under a warning-only promptIndependent evaluators report safety gains and shutdown-control failures in DeepSeek V4.1 Flash-1017. Cortist separates the benchmark from real-world risk.
86%Analytical choices matching the assigned agent pr…In a 50-analyst study, expert choices often followed a secretly assigned AI preference. The controlled setup shows why human review alone may not guarantee independence.
1,010Antibody variants built and tested in roughly ten…Anthropic reports laboratory improvements with biotech partners. Cortist checks what has actually been measured—and what remains untested in patients.
4 levelsWorkforce analysisA new task-based framework explains why job redesign, training and career mobility matter more than simple AI job-loss forecasts.
122Participants treatedIambic reports Phase 1/1b signals in carefully defined subgroups. No FDA approval or proof of superiority.
32 MINReported analysisA reported 32-minute GPT-Rosalind analysis yielded research hypotheses, not validated treatments.
AI AgentsIsolated microVMs · preview announcedAgent Compute Engine promises isolated microVMs, durable workspaces and identity-aware networking—but customers are still waiting for a preview.
AI RiskSafety analysis · new October 23 researchNew analyses from the UK AI Security Institute and Anthropic describe an AI-agent safety failure—and why stopping an unsafe plan proved difficult.
11 daysOpenAI-reported training pauseA September boundary failure led to an 11-day pause. OpenAI reports stronger containment, but independent validation is still needed.
+138.1%Bloom company revenue YoYQ3 revenue growth is real; future fuel-cell market estimates and gigawatt-scale contracts still need careful interpretation.
83%Retrospective diagnostic accuracyA Science study found striking results for sequential diagnosis. Real-world patient benefit remains unproven.
134Policies and regulationsAlmost 40% of 134 grid measures were introduced within a year. Who pays for power connections that are never used?
-3.7%US young high-exposure jobs, y/yADP and Stanford find pressure on young US workers in highly exposed occupations. The figures are real; their causes still need testing.
27% to 9%Share of junior technical hiresAnthropic's workforce study reveals a changing career ladder as AI agents proliferate. One company's experience is not the entire labor market.
64%Reported workplace AI useNearly 50,000 workers were surveyed. The divide between AI adoption and genuine learning opportunities may shape the future of work.
14.8BOpen DNA-language modelAn open 14.8B model brings DNA into a chat interface. Developer benchmarks are promising, but clinical validity is unproven.
3Supported answer typesA general-availability endpoint for bounded choices, probabilities and scores promises low latency and low costs for repetitive app workflows.
3 weeksIndependent pilotMETR and Redwood uncovered monitoring weaknesses that could conceal AI-agent actions. OpenAI says it patched the issues; outside verification is still missing.
10Voluntary incident practicesNIST, NSA and CISA set out voluntary safeguards for advanced AI agents. A US government label must not be confused with proof that ASI exists.
Google DeepMind's leader expects AGI within two to three years. That is an expert timeline, not a measured probability or a verified result.
1.0Distributed inference releaseIBM, Google Cloud and partners launch an open inference stack designed for agent workloads. The claimed savings are striking. The real test is the full cost of running it.
ScopeDefine the permitted taskMicrosoft puts enterprise agents inside its security perimeter. The harder question is whether a company can prove what each agent is allowed to do.
LatencyTime to finish one taskModal reports higher sandbox throughput with NVIDIA Vera. The result raises a practical question: how much agent time is spent waiting for ordinary compute?
BriefSet the audience and questionGoogle brings an editable planning moment into content generation. The opportunity is better direction; the responsibility to check the final story remains.
PyTorchExisting code, according to AmazonAmazon promises an easier path from PyTorch to Trainium. The real test is whether portable code becomes portable, cost-effective production.
5%Initial equity allocation proposedA proposed 5% equity commitment sounds substantial. What matters next is who controls the money, how it is spent, and what gets disclosed.
On-deviceModels running on a personal computerNew Windows laptops aim to make local models and personal agents practical. The deeper question is whether on-device AI becomes a daily habit.
120BSuper VL model parametersNew models span servers and local devices. For builders, deployability is starting to matter as much as raw benchmark scores.
5MSmallest model in the familyA new family of time-series models targets demand, sensors and operations—not conversation. It is a reminder that AI's next wave may be less visible.
4M5 Ultra Mac Studios in Apple's setupDistributed MLX inference makes a striking local demonstration. The next question is whether it makes economic sense beyond the lab.
InternalAccess for external evaluatorsSix METR staff are getting unusually deep access. The experiment could make frontier-model scrutiny more meaningful—if independence holds.
PlanDefine the intended taskGoogle is pitching its new model for complex coding and organizational tasks. The useful measure is not a demo—it is completed work without rescue.
IntentWhat the user wants to achieveOpenAI is moving from fixed chat replies to interactive responses assembled for the task. It changes how people might encounter software.
−75%Vendor-estimated average cost per taskAnthropic advertises a major reduction in average task cost. Builders should compare accepted outcomes, not just token price.
1TTotal model parametersA trillion-parameter preview raises the stakes for European AI. As of October 19, the weight release is still a promise, not a completed event.
Try another topic or search phrase.
Explore the AI & Jobs research hub ↗Explore AI & Robotics research ↗Explore AI & Cybersecurity research ↗Explore AI Chips & Compute research ↗Explore AI & Science research ↗Explore AI & Economy research ↗Explore AI & Governance research ↗We separate documented announcements from company claims and our interpretation. Every story links to its original source and shows what remains uncertain.
Why Cortist exists ↗Editorial accountability: Articles distinguish primary-source claims from independent verification. How our reporting is sourced and corrected ↗ · 25-article review record ↗
Explore Cortist's source-linked Reality Ledger: important AI claims, new evidence and open questions.
