AI is still improving exponentially on some tasks—but judgment-heavy work remains a major gap. METR’s updated time-horizon plot adds longer and more open-ended tasks to the evaluation suite behind its widely shared graph. The original analysis found that the task length frontier models could complete doubled every four months.
The update is less bullish on research automation. Models now outperform almost every human at software engineering, yet not one AI-designed evaluation was judged useful by METR employees. They also lagged humans at scoring philosophical arguments. Some newest models show “borderline above the bar” performance on evaluation design, up from “clearly below the bar” a year ago—but the tasks remain difficult to score and too few to support confident forecasts.
Why it matters: A four-month doubling rate on verifiable work should not be treated as a four-month doubling rate for autonomous R&D: METR found zero useful evaluation designs. For AI research teams, research taste and judgment—not just coding throughput—remain the harder automation problem.
HD Moore, Dragos’s chief research officer, planted nearly 2 million unique AWS keys, SSH passwords, private-repository references and tracking URLs into prompts sent to AI providers. Every request sent during several late-October test windows was stolen.
At some providers, attackers copied requests before a model ran. Parties tied to the breaches tested the stolen keys, searched S3 buckets, Secrets Managers and Parameter Stores, tried to enable models in AWS Bedrock, and added IAM users.
A Chinese AI lab fetched secret URLs from more than half of prompts sent to its models from July through October, including roughly two-thirds marked Zero Data Retention. U.S. hosts reselling those models triggered the same scraper.
Why it matters: OpenRouter, Vercel, and U.S. hosting customers cannot treat Zero Data Retention as a security boundary: Moore found requests stolen before inference, making on-premises deployment the most effective—but expensive—way to limit exposure.
AWS launched the Global Inference Gateway for SageMaker HyperPod, giving customers one endpoint to route inference requests across clusters in multiple Regions. The gateway evaluates each cluster’s capacity and load, routes requests to the best-suited option, and shifts traffic when a cluster fails.
A team serving four models across five clusters in two Regions previously managed 20 model-cluster endpoints; the earlier gateway reduced that to five, but clients still selected clusters themselves. The Global Gateway removes that routing work and the need to build cross-Region load balancers, DNS, and networking.
Requests stay private through AWS PrivateLink, with SigV4 and IAM authentication. Bagelcode says beta testing kept service running after one cluster failed.
Why it matters: Bagelcode’s small platform team can get cross-Region failover without building routing infrastructure; AWS now handles cluster scoring, health checks, and request forwarding.
MIT researchers built a machine-learning method that keeps models of chaotic systems within physically possible bounds, even over arbitrarily long forecast windows.
Instead of merely rewarding plausible behavior during training—or imposing arbitrary caps such as 100 degrees for temperature—the method has the model learn an energy-like function and a boundary for its dynamics. A projection layer then enforces energy dissipation as a hard constraint outside that region, pulling simulated trajectories back toward the bounded area.
Tests across textbook and highly complex multidimensional systems closely matched true outcomes, while conventional methods drifted into impossible predictions. The work appears in PNAS, and the team is making information about the approach freely available.
Why it matters: For weather forecasters and engineers modeling turbulent flows, MIT’s method offers a formal guarantee that simulations will not drift into physically impossible values over arbitrarily long forecast windows.
Anthropic’s IPO could channel as much as $156 billion into philanthropy, combining a public-benefit fund, employee share pledges, and company matching.
The PBF received 5% of Anthropic’s shares—nearly 850 million—giving it an estimated $80 billion endowment at a share price near $95. Its first commitment is $200 million for research on AI’s impact on workers. Employees have pledged shares, with Anthropic matching donations 3-to-1 for longer-tenured staff and 1-to-1 for newer hires.
The $156 billion total is an estimate: employees must honor pledges, share values can move, and donor-advised funds could delay deployment. Anthropic also gave San Francisco $50 million outside the PBF. Potential recipients include Longview Philanthropy, GiveWell, and AI-safety organizations.
Why it matters: If realized, Anthropic’s $156 billion philanthropic pipeline would equal roughly a quarter of all U.S. charitable giving in 2025. It also gives AI-safety organizations and EA-linked groups an unusually large potential funding source, though the final impact depends on employee pledges, stock prices, and payout timing.
AWS Quick now lets employees describe business processes in chat and turn them into reusable workflows. The update adds deterministic steps, longer-running workflows, large-data support, and multi-agent orchestration that combines specialized agents into one answer.
The redesigned web app puts daily briefings and pending work on the home screen, keeps workflows running after users close the tab, and brings desktop agents and skills to the browser. Quick connects to Meta ads and Amplitude, adds configurable content-safety controls, and supports Japanese, German, Simplified and Traditional Chinese, and Vietnamese.
AWS customer Pomelo says nearly four times as many employees now work with data each month, while one use case saves 80 hours monthly.
Why it matters: For Pomelo, Quick moves solution-building beyond the data team without expanding that team, while the fintech says its data specialists can focus on more complex work as monthly data users nearly quadruple.
AWS says distributed KV caching can cut self-hosted LLM inference costs by up to 40% and time to first token by up to 80%. The Agent Toolkit skill uses Amazon ElastiCache as a shared tier beside GPU fleets, letting servers reuse attention state across sessions, routing shifts, and server failures instead of recomputing long prompts.
The skill evaluates a workload before infrastructure changes, estimates savings with AWS’s KVTrace tool, and can configure existing vLLM or SGLang deployments. It works with Kiro, Claude Code, OpenAI Codex, and other coding agents that support Agent Skills. The figures are AWS benchmarks, not independently verified results; benefits depend on prompt length, repetition, routing, and cache cost.
Why it matters: For vLLM and SGLang teams, AWS's fit check makes ElastiCache a conditional expense: the skill stops when KVTrace estimates that savings will not cover the cache cost, before the team changes its infrastructure.