Lead Story · Industry

Amodei Answers a Critic With Nine Costly Safety Calls

Mythos-class models waited from February to June for cyber safeguards; distillation is “not a major threat” commercially.

Dario Amodei has answered criticism from @_NathanCalvin with a list of decisions he says were “very costly to us commercially while having strong safety benefits.” @kimmonismus picked out one: Anthropic held back broad access to Mythos-class models for four months to strengthen cyber safeguards. In Amodei’s words, Anthropic first had Mythos-class models in February but did not make them generally available until June, and offered them only to a limited set of partners focused on cyberdefense.

From posts by @kimmonismusRead the story

Agents & Coding

NVIDIA Puts Guardrails and Business Knowledge Around Agents

At GTC Berlin: the OpenShell Reference Architecture, and Agent Knowledge Foundation for Agent Toolkit.

AI agents need fast tool calls and security controls they can’t rewrite, NVIDIA AI says. Its OpenShell Reference Architecture combines the compute to run agents at scale with runtime controls over their access to tools and data.

From posts by @NVIDIAAIRead the story

Codex and Work Now Keep Prompt Caches Forever

The API gets 12-hour caching; @kimmonismus calls it a big upgrade for long coding sessions.

Tibo (@thsottiaux) says the API can now cache for 12h, and for logged-in users across Codex & Work the cache is kept forever. There is no penalty for coming back to an old conversation or resuming after a few hours, and users are “now guaranteed to hit the cache,” he writes.

From posts by @kimmonismusRead the story

Models & Hardware

NVIDIA Releases Flare, a Local Inference Server in Rust

Built for DGX Spark and RTX Blackwell GPUs; @MiaAI_lab calls inference engines the big trend.

NVIDIA just released Flare, a local inference server built for DGX Spark and RTX Blackwell GPUs, @MiaAI_lab reports. It is built for long chats, agents and tool calls, serves from Rust with no Python, plans memory up front and uses an FP8 KV cache by default.

From posts by @MiaAI_labRead the story

Latent Space Digs Into DeepSeek’s 96% Cache Hits

@swyx reads DeepSeek’s numbers as every GPU sold gaining ~30% in lifetime value.

Latent Space dug into DeepSeek’s production numbers in its AINews: 155T tokens a day from one cluster, with 96% input cache hits, and $38 revenue per GPU-hour at off-peak prices. At $5–8 per GPU-hour, it counts ~80% margin on GPU costs alone.

From posts by @swyxRead the story

@MiaAI_lab’s Ornith-1.5-122B Recipe Runs on One DGX Spark

256K context, a 1M total KV cache and 86 GB of memory, on TensorFold.

@MiaAI_lab published her recipe to run Ornith-1.5-122B on a DGX Spark, using her Ornith-1.5-122B-EXL3-EAQ 4 bpw quant. It runs on TensorFold with 256K context, a total 1M KV cache, 4 concurrent streams by default, 86 GB total memory and DFlash2 drafts.

From posts by @MiaAI_labRead the story

AI in Accounting & Tax

No AI accounting or tax dispatches from our sources in this edition. Accounting Today has not posted since Tuesday afternoon’s update.

Industry

NVIDIA Brings DSX POD and More Open Models to GTC Berlin

@kimmonismus says DSX POD turns months of data center build-out into days.

NVIDIA announced DSX POD at GTC Berlin. @kimmonismus calls it a bigger deal than most people realize: instead of taking months to build out a data center, compute capacity can scale in just a few days.

From posts by @kimmonismus, @NVIDIAAIRead the story

NHS England Expands Copilot to as Many as 500,000 Staff

Satya Nadella visits the Yorkshire Ambulance Service, which uses Copilot to cut administrative work.

Satya Nadella says NHS England is expanding Copilot to as many as 500,000 clinicians and support staff across the country. In the UK, he saw how the Yorkshire Ambulance Service is using Copilot to cut down on administrative work and make better use of its data.

From posts by @satyanadellaRead the story

The Wire

Quick Hits · Morning Desk

Statecharts, Then TLA+

@poteto suggests asking your agent to draw a statechart, then using a /create-formal-verification-skill made with pstack to find bugs through something like TLA+. You understand your code and the user flow better, he says, and fix issues users would otherwise see.

@poteto

Personal, Internal, Public

@mattpocockuk’s three steps to becoming a programmer: build personal software, then internal software, then public-facing software. His advice is to find even one tester who will use your app in real life. “AI changed NOTHING here.”

@mattpocockuk · @mattpocockuk

Determinism First

Asked whether skills should be deterministic scripts instead of markdown files, @mattpocockuk says: “Always choose determinism where determinism is an option. But it’s not always an option.”

@mattpocockuk

A Proofreading Prompt

@simonw says he uses AI for proofreading, and shares the prompt he uses, from his agentic engineering patterns guide.

@simonw

No-No Words

@swyx says Opus is great for writing “when you dont happen to have any random no-no words in your prompt.” His complaint: “give me prod, not god. i cant build an ai media org with this in the critical path.”

@swyx

The Breakout Talk

@swyx says @jediahkatz’s AI Engineer New York talk hit 100k views in 24hrs, an early candidate for the breakout talk of AIE NYC. It covers why smarter models let Cursor remove prompts and guardrails, and why Grok Bot is adding them back.

@swyx

Europe’s Factories

NVIDIA AI says Agile Robots, @Audi, @Dassault3DS, Schaeffler and @Siemens use its CUDA-X and Omniverse libraries, plus Cosmos and Nemotron open models, to test designs faster, simulate robots and improve factory operations.

@NVIDIAAI · @NVIDIAAI