Anthropic reconstructs NIST’s curve seeds

[Anthropic reports](https://www.anthropic.com/research/ecdsa-curve-origins) that an internal Claude model reconstructed NIST’s prime-curve seeds using a timestamp-seeded pseudorandom generator, and provides code to reproduce them. The result reduces uncertainty about their origin; the criteria for selecting candidate curves remain unresolved.

DeepSeek releases V4.1-Flash-1017 weights

DeepSeek · Hugging Face

[DeepSeek reports](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017) a Terminal-Bench 4.0 score of 42.3, up from 31.2, using the Claude Code harness. Its DeepSWE v1.1 score rises to 77.0 from 74.2 with Mini-SWE-V2. Both comparisons are vendor-reported.

Reflection releases Beam weights and post-training recipes

SiliconANGLE

Reflection makes Beam’s weights available alongside post-training recipes and technical resources for customization.

Microsoft details the training stack behind MAI-Base-2

Microsoft AI

Following yesterday’s MAI-Base-2 announcement, [Microsoft’s technical write-up](https://microsoft.ai/news/frontier-pre-training-at-microsoft-ai/) covers data strategy, compute efficiency, and the infrastructure behind pre-training. It also describes plans to bring internally trained models to Foundry.

OpenAI demonstrates dots handling an on-call investigation

OpenAI · 2m38s

The demo gives a persistent cloud agent a Slack identity and instructions to triage alerts. Its investigation opens in a Codex session that the engineer can steer from a phone or laptop.

Latent Space examines cloud-to-local agent state

Latent Space · 1h15m

The speakers argue for append-only event logs so coding agents can resume across environments, citing Pi’s persistence work. They also describe human review capacity as a constraint on agents that generate code overnight.

How Claude Code limits repeated review rounds

Anthropic · Claude developer blog

[The team describes](https://claude.dev/blog/solving-code-review-at-anthropic/) a review process that asks authors to reconsider their approach after three rounds and narrows later reviews to the diff. Humans retain responsibility for consequential decisions and changes to sensitive code.

Anthropic presents dynamic multi-agent workflows

AI Engineer · 1h41m

The talk demonstrates a lead agent creating and coordinating sub-agents as tasks change. Managed lifecycles and durable event logs support parallel execution and recovery without requiring developers to predefine every agent connection.

Anthropic reportedly offers Blackstone a yield above 10%

Financial Times via Techmeme

Following yesterday’s AI debt coverage, the FT reports that Anthropic is offering Blackstone a yield above 10% to help finance Google tensor processing unit purchases. The proposed debt package totals $60 billion, with Blackstone expected to invest $9 billion, according to the report’s sources.

Anthropic pitches rules for government release restrictions

Politico via Techmeme

Politico reports that Anthropic has circulated a draft AI safety bill to US lawmakers. The proposal would limit the circumstances in which the government could block a model’s release.

Replit’s Anthropic relationship is disputed

TechCrunch

[TechCrunch reports](https://techcrunch.com/2026/10/16/replit-is-no-longer-working-with-anthropic-ceo-says/) that Masad cites cost and performance as reasons for Replit leaving Anthropic; an unnamed source says the partnership remains in place.

Websites put restrictions on autonomous agents

The Verge

The Verge examines websites blocking agents as AI companies promote automated browsing and purchasing. Those restrictions affect whether an assistant can complete tasks across third-party services.

Workday agrees to acquire identity platform C1.ai

Portland Business Journal via Techmeme

C1.ai, formerly ConductorOne, manages identities for both people and AI agents. Workday has agreed to acquire the company, which has raised $111 million.