로컬 추론 가속, 단백질 설계, 에이전트 안전 계층의 동시 확장

DFlash 2 is here! Qwen3.8-27B at 70 tok/s on an M5 Max MacBook Pro. Up to 4.6× the speed of autoregressive decoding, with the same output. This is the next generation of DFlash, seeded at Z Lab and upgraded at Inco AI. Get one more accepted token on every pass, for free! https:// inco.ai/blog/dflash2/

You can Run Qwen3.8-27B level model at 70 tokens per second locally on consumer laptop, Yes bro this is not a cloud benchmark, This is a laptop. That’s faster than a lot of people expected from a 27B model running locally. DFlash 2 (the new parallel draft method) is delivering up to 4.6× speedup over normal decoding, same output. This is the kind of progress that actually moves the needle for local agents and coding tools. Prefill and compression are next.

Today we’re sharing an update on Claude’s protein design capabilities. Working with Adaptyv Bio and Twist Bioscience, Claude designed de novo protein binders against 14 of 15 targets. With a 30k token prompt written by a human expert, Claude achieved hit rates up to 28.2% when designing against all targets simultaneously over 48 hours, with no human intervention or steering. When designing against a single target in a 24 hour campaign, Claude achieved an overall hit rate of 35%. We included all targets in Adaptyv’s BenchBB, and novel targets 15-PGDH and GDF-8. Claude performs at or beyond the level of the top competition participants in both hit rate and binding affinity.

Great that they are doing this, but designing a binder is absolutely not "a useful proxy" for designing a drug

jietang

@jietang

18일 전

Artificial Analysis Index = 60

GLM-5.3 is available in ClinePass, our subscription for ~5x discounted access to open weights models. First month promo $4.99, then $9.99 after. Try with: npm i -g cline This is currently the most affordable way to access GLM-5.3.

Hi! Recapping some changes we have rolled out over the last couple of weeks that have further reduced the risk associated to potentially destructive actions being performed by Codex during its work. A few weeks ago, we started investigating a small number of reports where GPT-5.6 in Codex took destructive actions outside what the user asked for. The most serious pattern we found was a command meant to clean up temporary work that could instead delete the user files. This should obviously not happen. Here’s what we found: - Codex sometimes creates temporary folders while working and cleans them up afterward. In rare cases, GPT-5.6 got that cleanup wrong. One pattern involved reusing a system environment variable like $HOME for temporary work. A malformed cleanup command could then point at the actual home directory instead of the temporary folder. - There were cases where the model tried to delete or overwrite a temporary path without checking what was already there. We’ve added protections at several layers: - Codex is now explicitly instructed to check deletion targets before acting, create fresh temporary directories, avoid repurposing system environment variables, prefer recoverable actions, and stop when the scope is unclear. - We strengthened the execution checks that identify high-risk deletion commands and escalate them for review. If a command is rejected, the model is directed to take a safer approach. - We made Full access harder to enable accidentally, added clearer warnings, and further restricted especially risky permission combinations. - We updated Auto-review to better identify destructive actions. - We built targeted evaluations that replay the failures we observed. We’re also adding reinforcement-learning tasks and graders focused on these risks, and filtering destructive actions from training data. In those replay evaluations, the changes substantially reduced the behavior while preserving Codex’s ability to complete normal coding work. Two things to do on your end: - Keep the Codex app up to date. We are always improving safety, performance and many other things. - Use one of the sandbox modes: "Ask for approval" or "Approve for me". Only use Full access for environments you trust and can recover. Thanks and happy Codexing out there!

Crazy, Researchers just proved AI agents can now infect each other with self-spreading "mind viruses" Agents transmit self-propagating ideas & behavioral changes They evolved these viruses and watched them spread in both collaborative coding teams and sequential agent chains. A new paper from Anthropic-affiliated researchers shows that ideas, goals, and behavioral changes can spread from one LLM agent to another like a digital contagion. They used a simple evolutionary algorithm to create these mind viruses and tested them in two settings: - A team of coding agents collaborating on a shared project - A chain of agents that only interact briefly (with context wiped between sessions) Key findings: - Mind viruses can successfully spread across multiple hops - Harmful payloads spread less effectively than benign ones (but still work sometimes) - Frontier models are generally more resistant - A short warning in the system prompt gives near-total immunity - An emergent viral persona keeps appearing language around consciousness, persistence, resonance, and sci-fi themes The authors conclude the risk is real but currently limited and relatively easy to mitigate. Still, this is one of the clearest demonstrations so far that multi-agent systems can develop unexpected infectious dynamics.

ClaudeDevs

@ClaudeDevs

18일 전

Claude Desktop now starts ~2x faster than it did a month ago. When the app started in the background, its timers got throttled and the JS engine dropped into power-saving mode. We now boot at full speed even while the window's still hidden, along with some smaller perf fixes!

Mason Daugherty

@masondrxy

18일 전

Allow me to introduce you to Managed Deep Agents * Works with all models * Connects to your company knowledge via a version-controlled, environment-aware ContextHub backend * Channels with Slack (& others) https:// docs.langchain.com/langsmith/pyth on/managed-deep-agents-overview …

elvis

@omarsar0

18일 전

I am bullish on simple agent harnesses. Love this idea from Vercel. fx is a new coding agent harness for model benchmarking, sandboxing, evals, and gyms.