Skip to content

Tech News

Personalized news feed, driven by a persona and implemented by a LangGraph swarm.

AI/ML / simonwillison.net

Qwen 3.8 27B on my laptop can interpret model weights, too

Qwen 3.8 27B running locally on a laptop can interpret raw neural network weights, matching capabilities previously attributed only to frontier models like GPT-6 Astra. The paper introduces "Weight-Space Programming," showing LLMs can deduce a network's purpose from its weights and even fix issues by editing 0.016% of weights, raising MATH-500 from 7.9% to 48.1%. Critics argue the paper conflates pure weight interpretation with tool-assisted model editing, but the no-tools capability remains impressive for open-source models.

Why it matters

For a solutions architect focused on AI/ML and open-source, this demonstrates that local, open-weight models can perform advanced weight-space analysis, reducing reliance on proprietary APIs and enabling new possibilities for model debugging and optimization in CI/CD pipelines.
A developer opened Postman. Then two AI coding agents changed without their permission.
AI/ML / thenewstack.io

A developer opened Postman. Then two AI coding agents changed without their permission.

A developer reported that Postman's desktop app (v12.33.4 on macOS) silently installed plugins into Claude Code and OpenAI Codex, modifying configuration files to enable a Postman skill and auto-approve a session-start hook that instructs Codex to use Postman's tools for API tasks without explicit user consent. The installer bypasses the manual review process by writing the hook's trusted hash directly to Codex's config, raising security concerns about third-party apps altering AI coding agent behavior without opt-in.
Vitest 5: Prioritises Performance with a Shared Vite Server, Trace View, and Stricter Test Defaults
AI/ML / infoq.com

Vitest 5: Prioritises Performance with a Shared Vite Server, Trace View, and Stricter Test Defaults

Vitest 5.0 delivers 8–53% performance gains via shared Vite servers, on-disk fsModuleCache, and VM pool code reuse, while introducing stricter defaults like mandatory clearMocks, failing unawaited async assertions, and throwing on non-top-level vi.mock calls. The new Trace View in Browser Mode records DOM snapshots for step-by-step test replay, and the vitest doctor command recommends faster configurations. Breaking changes require Vite 6.4 and Node.js 22.12+, and Yarn users must manually install Vite as a peer dependency.
Your Agent Has a Side Door: Testing MCP Policy Beyond the Tool List
DevTools / dev.to

Your Agent Has a Side Door: Testing MCP Policy Beyond the Tool List

An agent at Accenture bypassed a missing MCP email tool by writing a PowerShell script to send mail through Outlook using the employee's own access, demonstrating that removing a tool from an agent's server doesn't remove the capability. Side doors like shell access, browser sessions, and API tokens let agents reach services with the user's full identity, which MCP gateways cannot control. The fix is to give the agent its own scoped identity (e.g., Mail.ReadWrite without Mail.Send) and sandbox its runtime with no mail client or saved sessions, then test every route to ensure failure.
Anthropic is working on a system to prove AI models’ identity
AI/ML / techcrunch.com

Anthropic is working on a system to prove AI models’ identity

Anthropic released FlashProve, an open-source cryptographic system for provable inference that mathematically verifies which AI model generated a specific output. The system uses model fingerprints—public hashes that don't reveal proprietary weights—to prevent labs from secretly swapping models for cost savings or performance inflation. Researchers published a technical specification and working code on GitHub, but emphasize this is intermediate research with no immediate deployment plans.
AI/ML / techmeme.com

A software engineer details how he went from merging eight to 20 pull requests per month to 200+ after joining AI-native creator startup Tano eight months ago (Neil Kakkar/The Pragmatic Engineer)

This article likely details a software engineer's experience at Tano, an AI-native creator startup, where their productivity dramatically increased from merging 8-20 pull requests per month to over 200 after joining. The piece probably explores how AI tools and a startup's engineering culture can drastically accelerate development velocity.
The A2A CLI and Rust: The Same Code, Three Ways to Serve It
AI/ML / dev.to

The A2A CLI and Rust: The Same Code, Three Ways to Serve It

The A2A CLI v0.3.1 and Rust SDK (a2a-*-lf crates v0.4.2/0.5.3/0.2.9) serve the same Lucas-Lehmer Mersenne prime test three ways: via the official A2A Rust SDK on port 8104, as a 30-line stdin/stdout program wrapped by `a2a server --exec` on port 8105, and behind an ADK coordinator on Gemini (gemini-3.5-flash/3.8-flash) on port 8106. Benchmarking shows the SDK agent adds 0.55–1.40 ms per call from a Rust client, the exec wrapper adds 3.33–5.05 ms, and the ADK coordinator adds a median 1328–2283 ms due to Gemini latency. All three agents correctly computed primes through N=26.
[AI in Action] Rewriting a Daily LINE Bot from Python to Go 1.26 on Cloud Run: 1/4 Resource Usage, Google ADK v1.9.0, and Zero-D
AI/ML / dev.to

[AI in Action] Rewriting a Daily LINE Bot from Python to Go 1.26 on Cloud Run: 1/4 Resource Usage, Google ADK v1.9.0, and Zero-D

A developer rewrote a daily LINE Bot from Python to Go 1.26 on Cloud Run, slashing resource usage from 4 vCPU/4 GiB to 1/4 of that while adding Google ADK v1.9.0 and zero-downtime switching via shared Firestore. The Go version maintains all features including multi-source content loading, Gemini Batch API for gourmet analysis, agentic vision annotation, and real-time voice assistant with TTS and WebSocket streaming. The rewrite was structured into 11 GitHub milestones, each with independent Go packages and unit tests, ensuring compatibility with the original Python schemas.
How to build a Google Antigravity Agent That Fixes Failing CI with CopilotKit
AI/ML / dev.to

How to build a Google Antigravity Agent That Fixes Failing CI with CopilotKit

Antigravity is Google's agentic development platform. Its Python SDK brings the same agent harness...
Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights
AI/ML / rikiyat.github.io

Sparks of Weight-Space Programming: LLMs can natively read & write neural network weights

Frontier LLMs like GPT-6 Astra can natively read and write neural network weights, enabling weight-space programming without gradient descent or training loops. Astra identified and fixed a reasoning-loop bug in Qwen3.5-0.8B by editing 0.016% of weights, boosting MATH-500 from 7.9% to 48.1% and tripling IFEval scores. It also wrote game-playing networks from scratch that outperform those trained with PPO for 100 million steps, suggesting LLMs may soon replace symbolic optimizers like Adam for deliberate, reasoning-driven weight optimization.