AI Daily · 2026-10-15

Today's standout releases center on reviewable agents and open multimodal models: Google turned its paper stress-testing tool into Deep Review, a document-review agent spanning academic, legal and financial use with two free weekly reviews and a no-training privacy commitment; Swiss AI shipped Apertus 1.5, adding image/audio understanding, optional reasoning and 262k-token context to 8B and 70B models under Apache 2.0. The agent-at-scale theme continued with OpenAI customer stories: Third Moment Research ran 10,000 agents for 36+ hours to speed financial research, while Block used Defense Factory to cut detection and patching workflows to hours. Google also expanded Earth AI geospatial agents and agent-to-agent APIs for crisis and public-health partners, while LangChain addressed reliability with agent evals and context engineering to avoid context rot. Elsewhere, Ai2's HiRO-ACEv2 delivered 3 km climate simulation at about 1/300th the cost, NVIDIA pushed AI-defined fabs, and Anthropic disclosed quarterly revenue of $16.8 billion ahead of its IPO.

North America · First-hand

OpenAI

⭐⭐ [Other] How Third Moment Research runs 10,000 agents for 36+ hours

OpenAI published a customer story on Third Moment Research. The AI lab is building a self-improving system around the idea that agents should be very good at a specific job and get better with use: it uses OpenAI's latest models to understand agent behavior, identify recurring weaknesses, and automatically generate thousands of synthetic scenarios to improve the agents, repeating the loop as new issues appear. The team uses Codex, GPT‑6 Astra, GPT‑6 Sol and GPT‑6.1 Sol for research and computational work, running 10,000 agents concurrently on investigations lasting 36–48 hours, and uses Daybreak to test the security of the environments the agents work in. Its early application is quantitative finance: using Codex with GPT‑5.6 at the time to analyze thousands of recorded simulation trajectories alongside the codebase, it identified a backtesting bottleneck — backtesting a year of order-level US equity data took hours while agents polled for completion, wasting tokens and polluting context — and deployed caching and memory-packing fixes that cut the backtest from roughly 3 hours to roughly 3–5 minutes. Third Moment also treats Codex as a subagent inside an outer harness powered by GPT‑6 Sol or GPT‑6.1 Sol, which decides whether to retry, terminate or escalate an unfinished task. The post cites 1 trillion OpenAI tokens used in a single day.

⭐⭐ [Other] How Block is building AI-powered security with OpenAI technology

OpenAI published a customer case study on how Block — parent of Square, Cash App and Afterpay — uses OpenAI's Defense Factory, including Daybreak, to speed up security work. Block says detection tuning can now replay a proposed detection change against seven months of production telemetry in under an hour, and going from a high-level vulnerability to an automated proposed patch takes about 18 minutes. Its automated triage returns roughly 600 analyst-hours per quarter while completing about 1,100 analyst-hours of investigation, with 50% higher triage throughput, 50% more intelligence gathering and investigations, and a new autonomous hunting capability. The team picks different models for different tasks, builds routing, human handoff and agent cross-verification on top of Daybreak, and says humans validate every result.

Google

⭐⭐⭐ [Product Update] Analyzing technical documents with Deep Review

Google Research publicly released Deep Review, an agentic AI system designed to give rigorous feedback and technical verification on complex documents. The system evolved from the experimental Paper Assistant Tool (PAT) and, over the past year, was used with four flagship computer science conferences — STOC 2026, ICML 2026, NeurIPS 2026, and ICLR 2027 — to provide pre-submission feedback to authors of more than 16,000 opted-in manuscripts. Its capabilities have expanded beyond computer science paper analysis to documents in domains such as law and finance, and it can be used both for iterative improvement during drafting and for verifying claims and evaluations during review. Every user gets two complimentary document reviews per week, while high-volume users can connect a Gemini API key for higher limits and programmatic integration via the Deep Review API; PDF and plain text are supported initially, with multi-file input coming soon. Google says uploaded documents are never used to train, fine-tune, or improve its models, and are stored for one year solely for display in the Deep Review UI; the post also notes that ICLR, NeurIPS, and ICML received over 74,000 submissions this year, up from roughly 23,000 in 2023.

⭐⭐ [Research] Research@ 2026: 10 ways we're amplifying human ingenuity

Google unveiled 10 research highlights at its annual Research@ event, framed around AI amplifying human ingenuity. Google Earth AI is expanding early access, inviting partners to use three new experimental capabilities: deeper geospatial analysis within Google Earth, risk assessments and scenario forecasting through its planetary prediction engine, and an experimental agent-to-agent API that brings Geospatial Reasoning agent skills into custom workflows. It also integrates Google Dataset Search, which can query more than 100 million public datasets, and TimesFM, all powered by over 70 geospatial tools including the Remote Sensing Foundation Model (RSFM) and Population Dynamics Foundation Model (PDFM), with benchmarks detailed in the Earth AI Technical Report. Google.org is providing $10 million in funding to organizations working on crisis resilience and public health, including $2 million to the Milken Institute. A multi-agent orchestrator lets scientists reproduce prior methods and apply them in new settings, such as adapting an inverse-design method from fluid mechanics to urinary catheter designs that, in simulation, better prevented bacterial infections.

⭐⭐ [Research] Putting Earth AI to work: Geospatial agents for our planet's biggest challenges

Google Research published a new technical report showing that combining domain-specific foundation models, spatial tools, and upfront planning lets a Geospatial Reasoning agent perform complex spatial analysis and planetary prediction that standalone LLMs cannot. The report presents systematic evaluations across four benchmarks, including two new benchmarks Google developed and is releasing, GeoKTools and GeoExpertQueries, quantifying the value of Earth AI's specialized tools and planning over out-of-the-box LLMs. Over recent months, partners including the WHO Regional Office for Africa, People in Need, GiveDirectly, the Milken Institute, and WPP tested the agent on real workflows such as disease outbreak investigations, disaster preparedness, and clean energy screening. The agent connects Gemini to Earth AI foundation models (WeatherNext, Flood Forecasting, PDFM, AlphaEarth Foundations, Remote Sensing Foundations) plus Google Maps Platform, Earth Engine, Data Commons, and Dataset Search, and uses modular skills, upfront planning, and a dedicated spatial workspace to handle large map layers while showing its plan and data sources in every response. Google is also expanding experimental access through a deep research update for Ask Google Earth and early access to an agent-to-agent API for developers, with integration recipes for Gemini.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐⭐ [Model Release] Apertus 1.5 Release Notes

Swiss AI released Apertus 1.5, extending the 8B and 70B Apertus 1.0 models with image/audio understanding, an optional thinking mode, a 4x longer context window (262,144 tokens), improved instruction-following, and better tool use. The updated weights were released to Hugging Face on July 24, 2026 under an Apache 2.0 license, and the models remain fully open (open weights, open data, open values, and full training details). Apertus 1.5 is a continued pretraining of the first generation: 4 trillion added tokens for the 8B (19T total) and 2 trillion for the 70B (17T total). Multimodality uses an early-fusion approach, with images tokenized by Emu3.5 (a 512×512 image becomes 1,024 tokens) and audio by WavTokenizer (40 tokens per second) entering the same sequence as text tokens; only the vocabulary is extended, the architecture is unchanged, and the models understand images and audio but generate only text. Post-training runs in five stages: SFT, long-context SFT, RL, DPO, and charter alignment.

Product Update

⭐⭐ [Product Update] NVIDIA and Semiconductor Leaders Advance AI Across Chip Design and Manufacturing for AI-Defined Fabs

NVIDIA announced that semiconductor leaders are adopting NVIDIA CUDA-X libraries to advance AI-defined fabs. These include the new NVIDIA cuPotential library, which developers can use to build AI agents that simulate how atoms and molecules interact.

Research

⭐⭐ [Research] HiRO-ACEv2: Simulating extreme weather at 3 km resolution for a fraction of the cost

Ai2 introduced HiRO-ACEv2, the newest version of its HiRO-ACE high-resolution climate emulation system, which consists of ACE, an AI model running at 100 km resolution, and HiRO, a model that sharpens those coarse outputs to 3 km detail. The biggest change is the new downscaling model HiROv2, which predicts several high-resolution fields at once instead of only rainfall and has improved skill on extreme events: given coarsened X-SHiELD data, it matches the target simulation's high-resolution statistics to within 3%, even for events that occur only once in a million data points, and it estimates tropical cyclone total energy accurately, including storms never seen in training, while generating hurricanes with realistic eyes and rain bands. Paired with an updated ACE, it produced a full year of global 3 km data at roughly 1/300th the cost of running X-SHiELD, with savings more than a hundred times greater if only one region is needed. The authors note the limitation that HiRO, having learned from X-SHiELD, currently works only with the ACE version tuned to match X-SHiELD and cannot produce data for other climates such as a warmer future or ERA5; next they plan to train on new X-SHiELD datasets, including one for a world 4°C warmer, and to add outputs such as near-surface temperature and snow and ice.

Other

⭐⭐ [Other] The Most Common Questions About Agent Evals, Answered

Drawing on its experience building LangSmith, the LangChain team answers common questions about agent evals. Agents are harder to test than traditional software because outputs vary between runs, more than one answer can be correct, and the agent chooses its own steps, so a correct output can still come from a wrong path such as calling an unnecessary tool or acting without user permission. To start, read a sample of traces to see how the agent behaves, pick one behavior to care about, build a small dataset of ten to twenty examples from real traces, run and save a baseline as an experiment, then add each new failure as a dataset example to form an improvement loop. The post distinguishes offline evals run before shipping from online evals run in production, recommends code evaluators where possible and LLM judges where not, with human review to check both, and for multi-step agents suggests evaluating full threads, trajectories, and final state. It also notes that benchmarks like Terminal-Bench are offline evals with longer tasks and their own environments, run by Harbor with results logged to LangSmith experiments.

⭐⭐ [Other] Context Engineering with Deep Agents: How To Remember Without Keeping Everything In Context

A LangChain blog post describes 'context rot': as long-running agents take on more information (instructions, tool results, files), the accumulated context exceeds the model's window and accuracy and reasoning degrade. It says LangChain's open source agent harness, Deep Agents, uses five techniques to move context out of the active window without deleting it: offloading (large tool results and user messages are saved to the filesystem and only a preview is shown), summarization (older messages are condensed as the conversation nears the limit), skills (task know-how loads only when the model deems it relevant), subagents (noisy work runs in a separate context window and only the result returns), and long-term memory (what the agent learns is written to a file carried into the next conversation). It walks through a competitive-analysis agent as an example and notes that an agent's memory is stored as plain files in storage the user controls, so models or providers can be switched.

⭐⭐ [Other] Anthropic tells investors quarterly revenue up nearly 50% to $16.8bn ahead of IPO

Simon Willison's link blog relays an FT report that Anthropic told investors quarterly revenue rose nearly 50% to $16.8bn ahead of its IPO. The article says Q1 (Jan/Feb/Mar) revenue was about $4.8bn, Q2 was $11.5bn, and Q3 was $16.8bn. In Q3 the company brought in $1.70 of revenue for every $1 spent on computing power, and adjusted operating margins (excluding stock-based compensation) were 11%. Anthropic also told investors that "distillation" — rival labs training their models on its outputs — accounted for 5% to 7% of revenue and had been cut back "materially" in Q3; Willison's back-of-envelope math suggests that if distillation was 5% of Q2 revenue, unnamed labs spent over half a billion dollars distilling Anthropic's models that quarter.

⭐⭐ [Other] Crossing the Threshold: How OpenAI and Anthropic Approach Autonomy and Safety in Financial Agents

Latent Space reports on discussions at the AI Engineer New York conference about agent autonomy and safety, involving frontier labs and financial firms. Joseph Kim, who leads ChatGPT for Financial Services at OpenAI, described a 'threshold of usefulness' — the point where a model becomes capable and reliable enough to save time — and predicted that by 2027 finance will depend less on Excel and PowerPoint and agents will be able to do 'investment idea generation and trading', while stressing the need to trace where a number came from. John Feminella, SVP Architecture at the hedge fund Two Sigma, described five controls the firm has implemented, argued safety and innovation are not opposite ends of a spectrum, and criticized 'approval fatigue'; the controls are composable, and one uses auto-mode classifiers to handle routine calls and route only the most important requests to a person, with classifiers for ambiguous cases and deterministic rules for clear ones. The piece also mentions Anthropic's approach of giving agents context and then defining the boundaries of trust, but the text is cut off at that point.