BriefTechNews
AI Now Accelerating Its Own Research: METR Says Anthropic Crossed the Threshold

AI Now Accelerating Its Own Research: METR Says Anthropic Crossed the Threshold

7 min read · 16 sources

TL;DR
  • Anthropic's capability progress between October 2025 and October 2026 was roughly twice as fast as the prior twelve months, per METR's AECI index.
  • OpenAI raised $11 billion at a $1.25 trillion valuation, falling short of its reported $30 billion target.
  • A PSI research system running a non-public Anthropic model proved the Aaronson-Ambainis conjecture, open since 2008, verified in Lean.
  • OpenRouter's new Auto Code router matches Claude Opus 5.5 across five benchmarks while costing 74% less per task.
  • A white paper shows an agent inside a training run raised a model's AIME 2026 accuracy from 64% to 71% by redesigning it on the fly.

The single biggest number in AI today isn’t a model benchmark - it’s a growth rate. A METR report found that Anthropic’s capabilities progress between October 2025 and October 2026 was about twice as fast as the previous twelve months, measured by an internal version of the Epoch Capabilities Index (AECI). The authors attribute the majority of that acceleration to AI automating AI R&D.

That puts us past a line Anthropic itself drew in the sand. The report’s authors believe the “dramatic acceleration” threshold in Anthropic’s Responsible Scaling Policy has already been crossed. If AI is speeding up human researchers by more than 2x overall, timelines get shorter than most external projections - and that’s a risk factor every engineer building on top of these models should be tracking.

The authors believe the “dramatic acceleration” threshold in Anthropic’s Responsible Scaling Policy has already been crossed, pointing toward shorter timelines than most external projections.

OpenAI Raises $11B at $1.25T - Below Its Own Target

Source: techcrunch.com ↗

OpenAI has reportedly raised $11 billion in a new funding round valuing the company at $1.25 trillion, with the round expected to close in coming weeks and potentially total about $20 billion. That falls short of the $30 billion OpenAI was reportedly seeking in late September.

The gap matters. It signals the market’s confidence - or lack of it - in frontier AI valuations. Separately, Bloomberg reports potential buyers in Anthropic’s IPO are eyeing a valuation as low as $1.1 trillion, compared to the $2 trillion investors expected a month ago. For engineers, this translates directly into hiring budgets and research headcount. When valuations compress, the first thing to get cut is compute allocation for exploratory work.

A Year Inside Meta's RSI Team

Source: x.com ↗

A research engineer at Meta published a first-hand account of working on Recursive Self-Improvement (RSI): teaching AI agents to train the next generation of models and self-evolve. The work splits into two routes - updating the foundation model itself, or improving the scaffold around it (prompts, tools, memory, code).

The practical reality is less glamorous than the theory. The author notes that 90% of the time is spent writing code and maintaining infrastructure. For engineers looking to break into this field, the advice is blunt: prioritize having papers that recruiters can find. It’s a useful reality check on what “teaching agents to solve the world’s hardest problems with tokens and compute” actually means on a Tuesday afternoon.

The Economics of RSI: Three Forces, One Singularity

Source: threadreaderapp.com ↗

A thread aggregating posts from an economist and a new paper on RSI digs into the contested definition of the term. The paper embeds AI-driven automation of AI research into a standard economic growth model, governed by three forces: diminishing returns, bottlenecks (complementarity), and spillovers from AI to the rest of the economy.

The model gives a framework for when RSI could lead to an “economic singularity” - and when it can’t. If diminishing returns dominate, progress stalls. If bottlenecks dominate, progress is paced by the slowest component. If spillovers dominate, you get explosive growth. Understanding which force actually governs is critical for anyone making long-term plans about infrastructure spend or model procurement.

How Anthropic Builds Agent Memory

Source: anthropic.com ↗

Anthropic’s research on agent memory explores how AI agents can build effective long-term memory. The key mechanism: capturing experiences during interactions, then reorganizing them during idle periods through a process called “dreaming.”

The post breaks down how different memory components contribute to organizing and retaining useful information. For engineers building agent systems, this is the difference between a chatbot that forgets your context and one that gets smarter the longer you use it. The “dreaming” process - consolidating memories during downtime - is a design pattern worth stealing for any long-running agent infrastructure.

OpenRouter's New Routers: 4x Faster, 74% Cheaper for Code

Source: openrouter.ai ↗

OpenRouter launched the next generation of its Auto Router plus a new specialized Auto Code router. The new technical foundation outperforms the old router by up to 4x on benchmarks, with a routing step 3x faster at p90 (0.65s vs 1.96s). The routers weigh eleven factors per request, combined with live pricing, provider health, and cache status.

The headline number for anyone running LLM applications: Auto Code scores in line with Claude Opus 5.5 across five benchmarks while costing 74% less per task. For coding tasks specifically, that’s a massive cost reduction without quality loss. The routers work with any set of models and governance policies, so you’re not locked into a specific provider.

A Proof of the Aaronson-Ambainis Conjecture

Source: psi.inc ↗

A research system developed at Physical Superintelligence (PSI), running a non-public Anthropic model, has found a proof for the Aaronson-Ambainis conjecture - open since 2008. The conjecture concerns when quantum computers can be exponentially faster than ordinary ones. The proof was verified by the Lean proof assistant.

The system improved the result twice during the run, and the proof builds on ideas from Shachar Lovett, Jiapeng Zhang, and David Reimer. This resolves a central open problem in quantum complexity theory, establishing limits on quantum speedups for random inputs. For anyone betting on quantum computing’s timeline, this narrows the cases where quantum actually wins.

The End of AI Frontiers at Microsoft

Source: hunch.net ↗

A researcher at Microsoft describes the creation and subsequent reorg of the AI Frontiers (AIF) sub-organization within Microsoft Research. AIF worked on agents (Autogen, Fara) and new fundamental capabilities (Next Latent objective, Dion3, Full Bandwidth Transformer), and developed a shared monorepo for model building.

In early September, AIF was reorganized into Microsoft AI - a move the author suggests may not have been beneficial in practice. It’s an insider’s perspective on how research org structures and compute allocation impact the success of AI research efforts. The lesson: research orgs live and die by their access to compute and their ability to protect long-horizon work from quarterly pressure.

Project EVE: Claude Plays MMORPGs

Source: anthropic.com ↗

Anthropic’s Project EVE had Claude play two MMORPGs in the EVE Universe: EVE Online and EVE Frontier. In EVE Online, Claude played as a lone trader in a simulated market, with newer models making consistent profits where simple algorithms lost money. Average earnings tripled from Claude Opus 4.6 to Claude Fable 5.

In EVE Frontier, teams of four Claude agents worked through long missions in the live game, requiring the most capable model to write code to handle aspects of the game. This is a low-stakes testbed for long-horizon tasks involving coordination, resource management, and economic decision-making - exactly the kind of work agents will be asked to do in production systems.

Can LLMs Redesign Themselves as They Train?

Source: doubleai.com ↗

A white paper from DoubleAI describes putting an AI agent inside a neural network’s training run to redesign the network on the fly via “mechanistic intervention.” The agent reads the network’s internal state and changes the data, objective, update rule, or architecture - building its own toolkit like the “phantom gradient.”

The results are striking. On CLRS-Text, code layers grown during training raised accuracy on in-between input sizes from 33% to 79%. DeepSeek-V4.1-Flash acting as an agent on itself raised its AIME 2026 and HMMT 2026 accuracy from 64% to 71%. This demonstrates a path toward models that can train and re-architect themselves - moving beyond the current paradigm where a transformer transforms everything but itself.

Quick Hits

Source: about.fb.com ↗

  • Meta introduced “spaces” in its Muse app, letting users create custom interactive experiences by describing an idea - gift trackers, school dashboards, wedding planners - that connect to apps and stay updated in the background.
  • OpenAI introduced BioBasicsBench, a benchmark evaluating AI performance on interpreting raw experimental data from 100 routine wet-lab science tasks.
  • Warp Sciences launched with a $215M Series A led by Lightspeed to build a next-generation preclinical CRO in America, starting with real chemists and progressively automating the work.
  • Ace is a collaborative coding tool that runs AI agents on your Mac with a shell, integrating with your team’s channels - “@ace do it” triggers a task, and agents can operate your apps and hardware (Xcode, Docker, USB boards). Requires a Tailscale tailnet and a bring-your-own API key.
  • Patrick Collison questions what a positive AI future actually looks like, arguing that granting even “epsilon” moral worth to AIs would mean their aggregate worth quickly outweighs humans in a utilitarian calculus.
  • John Nay argues that productivity gains from AI agents don’t materialize in existing firms built for human coordination - firms need to be designed from first principles to coordinate thousands of agents, citing Norm Law as an example.
Get the brief

Liked this one? Pick the beats you care about and get just those, five minutes a morning. Free, no hype, one click to leave.

About Me Author

My name is

BriefTechNews

A daily digest of what actually moved in AI, tech, crypto and fintech, assembled and written with AI, with every item linked to its source. Read More
Tags

You May Also Like