Philippe 🇪🇺

You can now enable Claude to use your computer to complete tasks. It opens your apps, navigates your browser, fills in spreadsheets—anything you'd do sitting at your desk. Research preview in Claude Cowork and Claude Code, macOS only.

@scaling01 I don’t think so. I am most worried about the GPU failure rate as they can’t be swapped out. And they are targeting a higher than terrestrial operating temperature.

okay actually, the biggest problem might actually be radiation

datacenters in space are silly 100kW isn't even enough to power a single GB200 NVL72 but sure let's spend 100 million just for launching the damn thing, while on earth you could buy like 30 GB200 NVL72 for that price

So it’s not one 100kW AI mini sat per Starship. A Starlink v3 launch will deploy 1MW of solar… AI sats target to pack up to 5x the power density and Starship v4 targets to double capacity. So for the full-size AI sats we are looking at 10x 1MW sats for a $20M launch a decade out Now that launch cost << sat cost, sat cost is the target. Massive solar efforts are underway. But Nvidia profit margin and TSMC availability are the blockers: here comes the Terafab project.

RT @Tesla: Announcing Terafab: the next step towards becoming a galactic civilization x.com/i/broadcasts/1…

Caught up with @karpathy for a new @NoPriorsPod: on the phase shift in engineering, AI psychosis, claws, AutoResearch, the opportunity for a SETI-at-Home like movement in AI, the model landscape, and second order effects 02:55 - What Capability Limits Remain? 06:15 - What Mastery of Coding Agents Looks Like 11:16 - Second Order Effects of Coding Agents 15:51 - Why AutoResearch 22:45 - Relevant Skills in the AI Era 28:25 - Model Speciation 32:30 - Collaboration Surfaces for Humans and AI 37:28 - Analysis of Jobs Market Data 48:25 - Open vs. Closed Source Models 53:51 - Autonomous Robotics and Atoms 1:00:59 - MicroGPT and Agentic Education 1:05:40 - End Thoughts

When I was on the ISS for my nearly year long mission, there was a telomere experiment comparing my telomeres to my earth baseline and my twin brother as a control. Hypothesis was they would get damaged and worse due to the environment. Turns out they got better. Initially NASA thought maybe it was due to exercise and diet. After I returned we learned JAXA had a telomere experiment on some small worms the same time I was there. Their telomeres got better too. Never saw the worms doing any exercise. After further study determined it was the radiation.

@mweinbach @jtregunna I love ultrawide (21:9) for split screen. But I also can’t read at high ppi without 200% software scale. And my eyes can only see so much area. So a curved 34” UWQHD for $250 is nice.

@jtregunna 21:9 52-inch is the move x.com/mweinbach/stat…

How do people work with only one display? It feels impossible to be productive What are you swapping screens every 10 seconds

@KyleHessling1 @TheAhmadOsman Both desktop builds can be $4K budget.

That may be! And I love my m4 pro for light inference but primarily everything else. The issue that Ahmad is pointing out is that for 5249, cheapest config with 128gb m5 max, you’d get a lot more long term compute value putting that money towards gpus. But honestly it is a great option for single user inference, if it’d run minimax m2.5 in a good quant at 30tps, you could genuinely not need anything else and wouldn’t need to buy another laptop or desktop for a long time!

Only MoEs should be used on DGX Sparks Unified memory is bandwidth constrained, MoEs help a lot because only a small subset of parameters is processed per token In practice, MoEs are the difference between triple digit tokens/sec under concurrent load & single digit tokens/sec

> Go create value for others and don’t worry about the returns. If you create more value than you consume, you are welcome in any well operating community.

@TheAhmadOsman @KyleHessling1 It shows the 32GB RTX 5090 and 128GB M5 Max are comparable inference machines : 60-65 tok/s 4-bit gen, expected comparable prefill depending on format support.

@KyleHessling1 Yeah there are trade-offs However, a 27B dense model being on-par with a 200B+ MoE should tell us something :) x.com/theahmadosman/…

@vchennai2 claude —ide

I've switched back to cursor from Claude Code Claude code makes it far too easy to lose codebase context since you're only looking at diffs For an application that requires precision that starts becoming a problem real fast Curious if anyone found a good way around this

Released today: /loop /loop is a powerful new way to schedule recurring tasks, for up to 3 days at a time eg. “/loop babysit all my PRs. Auto-fix build issues and when comments come in, use a worktree agent to fix them” eg. “/loop every morning use the Slack MCP to give me a summary of top posts I was tagged in” Let us know what you think!

Isaacman decided that vehicles to get us on the Moon are better than vehicles to enrich contractors. arstechnica.com/space/2026/03/…

@AryanA9019 @VadimYuryev @jimmyjames_tech And the RTX 5080 gets 0 at metal. You wouldn’t compare metal scores so why are you comparing opencl scores for.

@VadimYuryev @jimmyjames_tech Metal ≠ OpenCL scores. Apple's Metal is heavily optimized → M4 Max: ~187k Metal but only ~116k OpenCL. M5 Max 232k Metal ≈ 140–150k OpenCL at best. RTX 5080: 284k OpenCL. Not "almost matching". Just bad apples-to-oranges comparison.

Breaking: Apple’s M5 Max is almost as fast as the Nvidia RTX 5080. I expected more but I thought we would get 48 cores. Looks like we get the same 40 as before. Thanks @jimmyjames_tech for finding! browser.geekbench.com/v6/compute/593…

GPT-5.4 Thinking and GPT-5.4 Pro are rolling out now in ChatGPT. GPT-5.4 is also now available in the API and Codex. GPT-5.4 brings our advances in reasoning, coding, and agentic workflows into one frontier model.

We're introducing Cursor Automations to build always-on agents.

@saxushu @LastofLazarus Artemis III **will** be stacked not long after Artemis II. The people who don’t know “what to launch” are no longer in charge.

@FdsPhilippe @LastofLazarus Nope. The "roof" for SLS is the EUS. The walls are already there. And this 3 year gap is a blatant lie, Artemis III could be stacked right after Artemis II. However, we don't have lunar landers. What the cadence for if the mission elements aren't ready at all?

If the SLS were a Falcon 9, then Jared Isaacman's "fantastic" plan would look like this. I could agree with the LEO A3 mission (well, SpaceX aren't ready anyway), but the ‘Standard SLS’ is nothing more than sabotaging NASA’s own SHLV capability.

@saxushu @LastofLazarus The point is they are building the roof before the walls. They are going back to basics: launching the rocket regularly so that engineers want to work there, don’t quit after each launch, while relearning everything before each one. Which 2nd stage to use is a mean, not a target.

@LastofLazarus Irrelevant details. It is an explanation and the point is, both are quite undersized and limits the future possibilities.

Voice mode is rolling out now in Claude Code. It’s live for ~5% of users today, and will be ramping through the coming weeks. You'll see a note on the welcome screen once you have access. /voice to toggle it on!