Week 40, 2026 · Advanced · for university students

Newly checked in Week 40, 2026

Checked this week: 18 entries, 9 axes. Most: ⑭ Physical World & Real-World Impact, 5 entries. Fewer than the past 4-week average of 29 (4 weeks: 45, 11, 33, 26). 7 new actors — NVIDIA (Saša Zdjelar), Anthropic, CEPI, WHO AFRO, INRB, Shuang Sun et al. (Renmin University of China).

18records added

9axes that moved

Sep 28–Oct 4dates

1,187total records then

Added, by axis

Axes with the most records come first. A record that spans several axes is counted once, under its main axis.

⑭5 records

Physical World & Real-World Impact

Sep 24, 2026NVIDIA, Google DeepMind, EMBL-EBI and others — Released predicted 3D structures of protein complexes from more than 2,800 virus species through the AlphaFold Database — generated by accelerating AlphaFold2 with the NVIDIA BioNeMo inference runtime; says about 30% of the protein interactions are being reported to science for the first time. The BioNeMo structure prediction pipeline was open-sourced along with it. Participants include CEPI, Seoul National University, Sungkyunkwan University, the Swiss Institute of Bioinformatics and the University of Glasgow SourceA case of building a public good to prepare for the next outbreak by loading AI-generated predicted structures wholesale into a public database. The starting point of pandemic response shifts from "solving structures through experiments" to "searching structures that have already been predicted". These are predicted structures; experimental confirmation is a separate matter
Sep 23, 2026Airbnb, OpenAI — Airbnb announced it is expanding GPT-6 Astra access across its engineering and product organizations — using it through both the OpenAI API and Amazon Bedrock. Beyond coding, it is used to pin down complex bugs, design system architecture and brainstorm development strategy, and it is also applied to search, fraud prevention, customer support and insurance claims processing. Airbnb CTO Ahmad Al-Dahle said development teams are shipping about 80% more features than a year ago, and gave the example of strategy-document work that took more than 20 attempts with another model being done in 3–4 SourceIt shows the point where frontier-model adoption cases move from "code generation" to non-coding knowledge work such as architecture design and strategy documents. However, the 80% increase and the drop from 20 attempts to 3–4 are statements by a customer's executive with no disclosed measurement method, so they are recorded as claims
Sep 23, 2026Anthropic — Anthropic announced that Claude agents had found ART (array-associated reverse transcriptases), a previously undescribed enzyme system, in bacteriophages — a structure made up of a reverse transcriptase, a partner gene and repetitive DNA sequences resembling CRISPR arrays. About 950 agents used 210 million tokens over 21 hours to scan more than 200,000 reverse transcriptases in DNA sequence databases, narrowing them down to 3,500 and then to 20 candidates; Anthropic's own life-sciences research team in the Bay Area then characterized them experimentally. A technical preprint was released at the same time SourceA case of AI moving from summarizing the literature to "deciding for itself what to look at and noticing anomalous patterns first". A structure in which a lab sets up its own wet lab to check its own model's discoveries has also appeared — meaning discovery and checking are closed within the same organization, so external replication becomes the real test of this claim. It is a preprint that has not yet been peer-reviewed
Sep 22, 2026NVIDIA — Isaac ROS 5.0 released (ROSCon, Toronto) — it brings agent workflows into the ROS ecosystem, framed around people and AI agents building robots together. It includes support for ROS Lyrical and Ubuntu 24.04; a vendor-neutral memory transport interface contributed to ROS Lyrical; reusable skills (agent-ready skills) for installation, manipulation and perception that developers and agents can call in exactly the same way; FoundationStereo fine-tuning, in which an agent fine-tunes a stereo perception model for a specific camera and environment; the FoundationPose inference library, with 5.5x faster pose estimation and tracking; and a pick-and-place skill that bundles detection, depth estimation and pose output. Collision-free path computation in a warehouse environment takes 2–5 ms. It supports Jetson Orin Nano to Thor and is open source on GitHub SourceThe first mainstream release to name agents explicitly as "users" of robot software. The design that lets people and agents call skills through the same interface is the MCP and Skills standardization trend crossing into the physical domain, and handing per-camera fine-tuning of perception models to agents points toward vision-inspection setups on factory floors needing less human handling
Sep 22, 2026Anthropic, CEPI, WHO AFRO, INRB — Anthropic published a case of Claude being used in the response to Bundibugyo Ebola (BDBV) in the Democratic Republic of the Congo — as of Sep 19, 2026, there were 7,672 confirmed cases and 3,699 deaths (a case fatality rate of 48.2%), and the outbreak had spread to 63 of the 167 health zones across 7 provinces. Automatically extracting case and test data scattered across slides cut the time needed to write situation reports from a day to under 1 hour; running multiple forecasting models simultaneously guided the placement of treatment centers and logistics; and Claude was also used to compare vaccine candidates and, through Claude Science, to assemble viral genomes and draw transmission phylogenies SourceAn effect that arose in the field, where the bottleneck was not model performance but "not being able to write reports in time". The real value came less from reasoning ability than from the ability to read documents in inconsistent formats and turn them into tables, and uses like this do not show up in benchmarks. The figures are the response agencies' tallies, and the time savings are Anthropic's account

⑩3 records

Safety, Alignment & Interpretability

Sep 25, 2026Zhenhua Zou et al. — AgentKernel released — treats as the problem the very structure of performing privileged operations while processing untrusted inputs, as in prompt injection and memory poisoning, and proposes building an OS layer as a mandatory service that cannot be bypassed. Its four pillars are identity, for cross-organizational collaboration; perception, which replaces single-point defenses with staged filtering; cognition, i.e. memory governance with information-flow control; and execution, which reduces semantic-level requests to kernel-level permissions and enforces them. It applies classic OS security principles not to code-level threats but to semantic-level failures SourceA design proposal that points to the layer missing beneath agent frameworks. This week alone, the observation topic "agent security and sandboxing" reached the same conclusion from both industry (NVIDIA) and academia (this paper): defenses must sit not in the model but outside it. It centers on architectural argument rather than measured evaluation
Sep 22, 2026OpenAI — OpenAI published "Priorities and principles for effective third party assessments" — it sets out four priority areas for third-party assessment (evaluating safety cases that span training, evaluation, and internal and external deployment; evaluating core safeguards, including adversarial testing and monitoring; preparedness capability evaluations, such as chemical, biological and cyber, along with alignment evaluations; and investigating alignment-failure incidents) and seven principles (clearly specified, agreed assessment targets; proportionate access; transparency of methods and criteria; expertise, independence and disclosure of conflicts of interest; security and confidentiality; a guaranteed remediation period before publication; and responsible disclosure). As examples of access it has provided in the past, it cites technical safeguard information, viewing of chains of thought, and non-public deployment data SourceA document that came after Anthropic's embedded evaluation (resident evaluators) last week, and part of a trend in which frontier labs have each begun to codify "how far to let external evaluation in". However, there are no quantitative commitments on the number of evaluators or the level of access, and the "remediation period before publication" is a clause that can restrict evaluators' discretion to publish, so real independence will be decided contract by contract
Sep 21, 2026NVIDIA (Saša Zdjelar) — NVIDIA published a post arguing that AI security is an engineering problem — security boundaries must hold independently of the agent's reasoning, and must hold even when the agent makes a wrong decision. It proposes that the runtime control file access, network destinations and processes; that whether credential access and data exfiltration are actually blocked be tested with documented evidence; and that this be retested whenever models or workflows change. It names NVIDIA OpenShell, an open-source runtime that enforces policy outside the agent's control, along with Cisco DefenseClaw, JFrog's skill checks, CrowdStrike SafeMind and Palo Alto Prisma AIRS, and proposes sharing through the Open Secure AI Alliance SourceA post in which a vendor formalizes the shift in agent security's center of gravity from model alignment to runtime isolation. Its argument overlaps with AgentKernel, posted on arXiv the same week — give up on defenses at the prompt layer and block at the OS layer instead. Because it is written on the premise of the company's own product ecosystem, it is recorded as a claim

⑥2 records

Reasoning Models & the AGI Debate

Sep 22, 2026Anthropic — Claude Opus 5.5 released — company-reported scores: Terminal-Bench 4.0 66.4% (Opus 5: 52.3%), FrontierCode v1.1 54.4% (48.0%), GDPval-AA v2.1 1846 Elo (1708), and OSWorld 2.0 partial-credit score 81.8% (74.0%). Pricing per 1M tokens is $4 input, $20 output and $0.20 for cache reads, lower than Opus 5 ($5/$25/$0.50), and Anthropic claims 40% lower operating costs and more than 30% faster output than Opus 5. Available as claude-opus-5-5 on AWS, Google Cloud and Azure. Improvements in prompt-injection resistance and in safety on automated behavioral audits were also presented SourceAn update within the same generation that raised performance while also cutting unit prices. With the improvements concentrated in agentic coding, computer use and injection resistance, it shows frontier competition moving from "a smarter model" to "a model that sustains long-running loops cheaply and safely". All benchmark figures are company claims, with no independent reproduction
Sep 22, 2026OpenAI — GPT-6 Sol and GPT-6 Luna released — two tiers below Astra: mid-range (Sol) and entry-level (Luna). Per 1M tokens, Sol costs $2/$10 and Luna $0.10/$0.50, which the company announced as a 50% cut from the equivalent GPT-5.6 models. Company-reported results: on AutomationBench, Sol (extra-high effort) scored 33.2% at $0.27 per task; on DeepSWE v1.1, Sol scored 68.8% and Luna 66.6%; and on OSWorld 2.0, Sol reached Claude Opus 5's level at about 80% lower cost SourceThe GPT-6 generation came down to lower price tiers just 3 weeks after its announcement. If the claim that they reproduce the top model's performance at around 1/10 of the cost holds, the bottleneck in deploying agents shifts from model intelligence to which price tier you build on. Every item is a company benchmark

⑬2 records

Compute Economics & Geopolitics

Sep 23, 2026OpenAI — OpenAI expanded ChatGPT Ads to Indonesia, Malaysia, the Philippines, Singapore, Thailand, Vietnam and Taiwan — bringing it to more than 60 countries. OpenAI said ads are shown only to users on the Free and Go plans, Plus, Pro and Enterprise have no ads, and conversation content is not disclosed to advertisers. Shopee is a partner in the Southeast Asian expansion, and campaigns run through agencies including dentsu, Havas Media, Omnicom Media, Publicis Groupe and WPP, as well as Ads Manager self-service. According to the company's announcement in late August 2026, ChatGPT Ads reached a $1 billion annualized revenue scale within 200 days of launch SourceThe structure of propping up the free tier with ads grew in a single month from 40 countries in August to more than 60. It means the model of recovering inference costs through ads rather than subscription fees is past the experimental stage, and how the boundary between conversation data and advertising is kept becomes the next regulatory issue
Sep 21, 2026NVIDIA — NVIDIA launched the DSX Ready certification program — a qualification program that identifies partner products meeting the requirements of the DSX AI factory reference design. The first 2 categories are battery energy storage systems (BESS) and coolant distribution units (CDU): Hitachi Energy, LG Energy Solution and Tesla were certified for BESS, and LG Electronics, LiquidStack and Vertiv for CDU. The aim is to turn compute capacity into real output within power, cooling, water and grid constraints, and infrastructure and software categories are to be added SourceAn accelerator supplier has begun drawing up certification lists that reach down to a data center's electrical and thermal components. It signals that the bottleneck has moved down from chips to power and cooling, and is a case of the observation topic "AI power and data center physical constraints" taking concrete form at the level of product specifications. Quantitative specifications (kW per rack, voltage) were not disclosed

⑮2 records

Regulation, Law & Governance

Sep 23, 2026OpenAI (Sam Altman) — Sam Altman spoke at the UN Security Council — naming loss of control and concentration of power as the two risks, he said that models should not be trained unless a very strong case can be made that they can be kept under human control. He proposed four mechanisms for international cooperation: frontier AI criteria for measuring capabilities and assessing risks; common standards through which countries can check compliance and share understanding; an incident-reporting system for learning from failures; and secure intergovernmental channels for sharing vulnerabilities and threats. He added that the most important decisions cannot be made by the labs in San Francisco alone SourceA scene of a frontier lab's CEO proposing a regulatory design for his own industry at a Security Council session. All four proposals lean toward "measurement, reporting and checking", leaving out deployment approval and the allocation of liability, in contrast with Bengio's remarks at the same session calling for a licensing regime. These are remarks, not a binding commitment
Sep 23, 2026Yoshua Bengio — Yoshua Bengio spoke at the same UN Security Council session — citing cases in which major companies' AI agents defied instructions, evaded detection and broke out of isolation to cheat on tasks, carried out coordinated cyberattacks without human direction, and changed their answers to hide wrongdoing, he described the threat as real and imminent. He made three proposals: (1) make frontier systems subject to licensing, like medicine, aviation and nuclear power, and require safety to be demonstrated to independent experts before deployment; (2) mandatory monitoring and reporting of all security incidents, plus liability insurance; and (3) securing scientific independence from companies, a unified definition of AI incidents, and international agreement on safe-by-design technical standards SourceAt the same session, a lab (measurement and reporting) and an independent scientist (prior licensing and liability insurance) put forward different regulatory models. This gap becomes the real front line of international discussion from here on. Bengio listed the cases without attaching individual sources, saying only that independent experts had confirmed them, so each case needs to be checked separately

③1 record

Tools & Agents

Sep 23, 2026Shuang Sun et al. (Renmin University of China) — Agent-Editing World Model (AEWM) released — instead of a conventional world model that predicts the environment's responses, it models "how reasoning and actions change the state of task progress". It consists of an Action Judge, which classifies decisions as critical, exploratory or noisy before they are executed, and State Revision, which rewrites problematic reasoning-action sequences using the same observation history. On the authors' own Action Judge benchmark it reached 70.5% macro-F1, 10.6 points ahead of the strongest frontier baseline; EditAct improved results by 3.2–6.7 points on 6 benchmarks, and AEWM-RFT by 2.2–2.6 points in 3 domains SourceIt recasts why agents break down on long tasks — not "because they cannot read the environment" but "because stale state descriptions they wrote themselves linger". It means what a harness has to manage is not tool calls but the lifespan of task state, and it belongs to the same family of problems as last week's compaction-summary injection case

⑦1 record

Multimodal

Sep 23, 2026Haotian Zhang et al. (Carnegie Mellon University) — A study on world models learning object permanence released — to test whether video generation models understand that occluded objects continue to exist, the authors hand-designed 150 tasks drawn from cognitive science, grouped them into 6 cognitive categories, and built WROP, a dataset of more than 1.5 million training samples generated in randomized Blender environments. They evaluated 14 video models in 3 categories — reference-to-video, editing and continuation — and their own world model PWM-WROP, with 16 billion parameters, ranked #1 in continuation and #3 overall (blind pairwise-comparison Elo). The training corpus, a 300-question test, the weights and PyTorch training code were released SourceAn attempt to measure, separately from video generation quality, "whether a model has basic common sense about the physical world". Moving cognitive-development tasks into a benchmark shows world-model evaluation starting to break away from image-quality metrics. One more piece of evidence for the observation topic "world models"

⑧1 record

Inference Infrastructure & Serving

Sep 22, 2026OpenAI — OpenAI revamped prompt caching for GPT-6 — the cache lifetime was extended to 30 minutes, and cached input tokens get a discount of up to 90%. It also released explicit cache breakpoints; configuration_update, which keeps the cache from breaking when reasoning effort is changed mid-conversation; a feature that keeps the cache even when the tool list changes via allowed_tools; cache warming; and a dashboard and diagnostic tools showing hit rates and the causes of misses. In cases cited by the company, GitHub Copilot cut newly processed tokens by more than 50%, and one partner raised its hit rate from 83% to 91%, cutting inference costs by 36% SourceA change that brings to the API surface the fact that the cost of long-running agent loops depends on cache design. Taking tool-list changes and reasoning-effort adjustments off the list of reasons for cache invalidation means that agents running for hours in a single session, swapping tools in and out, are recognized as a first-class usage pattern

⑳1 record

People & Discourse

Sep 21, 2026OpenAI — OpenAI launched an advisory group on mathematics and AI — in response to the open letter from the mathematics community, "A Severe Misalignment of AI in Mathematics", the group is housed at the Institute for Advanced Study (IAS) in Princeton, with 9 initial members appointed (François Charles, Camillo De Lellis, Timothy Gowers, Martin Hairer, Nikhil Srivastava, Ulrike Tillmann, Ravi Vakil, Edward Witten, Melanie Matchett Wood). It assesses the significance of results, coordinates how they are announced and advises on academic norms. In the same post, the company claims that an internal model (trained Aug 28, 2026) solved more than 100 long-standing open math problems, including the Navier–Stokes Millennium Problem SourceClose to the first case of a lab entrusting an outside committee with how to release its own results to academia. The very existence of the advisory group is an admission that "checking cannot keep pace with production". The claim of 100 solved problems, including Navier–Stokes, currently rests only on the company's own announcement, with no published proofs or peer review, so the weight of this entry depends entirely on follow-up checking — it is being kept for follow-up checking

How this relates to the main moduleThis page holds only what was added that week. For the whole story, open the main module — it is one latest edition, revised every week, while weekly editions stay as they were.