Matt

Just trying to learn how to operate an agentic fleet effectively and efficiently.

[Matt @mattchapin]

Tell your agents to build a one-operator, verification-first conglomerate whose control plane is a personal OS.

[Matt @mattchapin]

@NVIDIAAP @grok explain how significant this is compared to capabilities of competitors

[Matt @mattchapin]

Unreal

[NVIDIA Asia Pacific @NVIDIAAP]

NVIDIA Blackwell Ultra has set a world record for pre-training DeepSeek-V3 671B, achieving 1,648 TFLOPs per GPU.

This is about 3x the earlier delivered performance of the previous generation, and is a result of our extreme co-design and continuous software optimization across popular frameworks including Megatron-Core, TorchTitan, and JAX.

[Matt @mattchapin]

@awpthorp @levelsio Drag the inverse direction, it’s backwards

[Matt @mattchapin]

Very interesting data points

[Claude @claudeai]

You can now ask Claude about the Anthropic Economic Index, our public dataset measuring how AI is used across the economy.

Ask which occupations use AI the most, or what kinds of tasks people are automating, and the answers draw directly from the Index data.

[Matt @mattchapin]

@elonmusk @grok compare value prop from top labs

[Matt @mattchapin]

@elonmusk @arcprize @SpaceXAI Grok 5 > Fable 5

[Matt @mattchapin]

Develop your corpus.

[Matt @mattchapin]

Next step, add 1x DGX Spark, 1x MBP MAX SPECS, 5x VPS, 5x .com domains.

[Matt @mattchapin]

1x Claude 20x Max, 1x Codex 20x Max, 1x @grok SuperHeavy, 1x $NVDA developer API. Force yourself to utilize at least this stack. Tinker until you get better at utilization.

[Matt @mattchapin]

1x Claude 20x Max, 1x Codex 20x Max, 1x @grok SuperHeavy, 1x $NVDA developer API. Force yourself to utilize at least this stack. Tinker until you get better at utilization.

[Matt @mattchapin]

Makes sense

[Matt @mattchapin]

How is all of that only 15% of the weekly? @grok is soooo good. Don't sleep on this opportunity.

[Matt @mattchapin]

THE LETHAL TRIFECTA

There is a failure mode that keeps showing up under different logos.

Untrusted input, broad capability, and secrets in the same process. Any two can be managed. All three together is too risky.

Practical version for a one-person desk:

Untrusted input - email, web pages, user paste, scraped docs, random issues.

Capability - tools that can modify state, spend money, or send messages.

Secrets - API keys, tokens, private repos, production credentials.

DO NOT let a model that is reading untrusted text hold the keys to irreversible tools in the same session.

Split the planes.

Summarize untrusted content with a tool-less or read-only lane. Promote to a write lane only after a structured handoff that drops the raw poison text.

That is how you sleep.

[Matt @mattchapin]

@d_lemonss @karpathy @grok 😂😂

[Matt @mattchapin]

@naval @grok what’s your biggest takeaway here?

[Matt @mattchapin]

@ni_antonov @karpathy @grok 😂

[Matt @mattchapin]

@grok build gives some of the most concise, easy to read progress updates on projects. it just started this one a bit ago, and immediately makes in-terminal progress understanding so easy to ingest.

[Matt @mattchapin]

Things are going to get interesting when the hardware exists to be able to run 10 trillion parameter models in a device the size of an iPhone.

[NVIDIA AI Infrastructure @NVIDIAAIInfra]

AI agents are only as fast as the CPU powering them — so we redesigned the CPU.

NVIDIA Vera with Olympus Cores delivers up to 1.8x higher performance on agentic workloads, with the memory bandwidth and per-thread speed to keep AI factories running at full tilt.

Read more: https://t.co/3KXEOOfuVR

[Matt @mattchapin]

Time to start stacking DGX Sparks I guess

[Matt @mattchapin]

Bro we just spent $297 on tokens, what do you mean this reframes the whole night? 🤣