Marshall Doyle

@emm0sh Sometimes DFM should be “design furthering manufacturing”. No need for novel manufacturing technologies if novel designs don’t require it. First create the need, then create the solution.

who wants to explain to the AI companies what DFM is?

@julien_c Gemma 4 by a country mile. I’m get higher throughput, better results on personal benchmarks, and most importantly I trust Google as a model supplier which gives me the comfort I need to run it as an agent with my data.

so…. Qwen3.5 or Gemma 4?

@LottoLabs Qwen 3.5 isn’t remotely as good as Gemma 4

Lots of people outing themselves by not knowing qwen3.5 27b is arguably better and been out for 4-5 weeks already

Gemma 4 is the best open source model ever released, fine tune coming soon...

any ai startup today

@LottoLabs I’ve been riding this wave since Llama, trained my own Alpaca models etc. Now learning MLX to do recreate the triton work I’ve done for a different ecosystem

I think the die hard local llm guys are kinda funny because at the end of the day your skills make you the best infra-guys at big labs

@Dylan_Morri @codyaims Kind of wild to suggest that milling machines have become ambiguously known as Bridgeports

@codyaims BRIDGEPORT

Reaching lvl 99 in Manufacturing is learning all of the “this is a company name, not the name of that tool / material”

@Zsshore @ItsElanaGold @hermeuscorp I thought Hermeus was in Atlanta?

@ItsElanaGold Don’t forget @hermeuscorp

@TheAhmadOsman Way back in the day I was fine tuning models on nasa documentation and got early models to very accurately describe complex systems and subjects. I distinctly remember showing a friend and realizing we would both be automated in our lifetimes.

Anyone remembers running Dolphin finetunes back in 2023 and feeling the AGI with 4096 and 8192 context windows? local LLMs have come so far

@aegrabs @nTopology is the best CAD software for this use case. Fully compliant Boolean logic based off of user scans and easily modifiable for manufacturing. Parametric is still great, but field based modeling seems to be where things are going

parametric design is so cool like wdym i can adapt my prosthetic arm to anyone by just changing a number

@UnslothAI @danielhanchen Unsloth, the best in the business!

We created a repo with 250+ notebooks for LLM training. Train locally on your device with 3GB VRAM or free on Colab. Learn the entire fine-tuning and inference workflow. Supports RL, vision, audio, embedding, TTS models GitHub: github.com/unslothai/note…

@aphysicist Maybe dm me a copy for peer review before publishing…

should i drop this lol

@RayaneRachid_ This looks incredible. I’ve had a ton of fun using three js with Claude code, can i ask where you got the 3D models from?

J'ai aucun mot.. Je viens de vibecoder un simulateur de combat aérien en 2 semaines, tout tourne dans le navigateur Vous pouvez test ici : ciwsimulator.com La stack : • React Three Fiber en base, wrapper React autour de Three.js pour gérer la scène. • drei pour tous les helpers (useGLTF, Sky, Environment etc). • Rendu WebGPU pour la perf, PAS du WebGL. • Tous les effets visuels faits avec TSL (Three Shading Language), shaders node-based, zéro GLSL. • Le tout en TypeScript J'ai utilisé GPT5.4 xHigh pour quasi tout, j'ai testé Opus mais c'était trop bugé L'afterburner, les balles, la haze, le bloom, la map, TOUT sauf le Rafale et le CIWS a été généré par l'IA N'importe qui peut build son propre jeu maintenant !

@iam_smx *trillioniare

Just found out that the shape store has a corporate office

@okthisisnagy Is this not still under defined? Much better, but critical dimensions missing?

For reference, this is what it should actually look like:

@Whateverface381 Great job! Do you know what center you’ll be going to?

I gotta control my language now

When you have a hammer, everything looks like a nail. When you’re a VC, everything looks like a capital problem. In the early days of Valar, a lot of VC’s passed on us because they didn’t believe we could raise the money. In one way, they were totally right. We raised less than a lot of our compatriots. But in a much more fundamental way, they were very wrong. You see, capital is not the only advantage a team can have. A team that is moving twice as fast needs half the operating capital per milestone. A team with intimate knowledge of the industry can spend a third of the CAPEX to get to the same place. Weirdly, I believe that this dynamic is more true in hard tech than software. A lot of software dollars end up going to sales and advertising, which is a really tough space to innovate on. You may occasionally see breakout successes with teams who know how to work the channels of earned media and vitality, but it rarely ever passes out of a normal band of acquisition cost. In this lens, the market capture advantage of having an extra $200 million in the bank begins to overshadow everything else: the details of the product, the quality of the team, etc., especially as software gets increasingly easy to build. I believe this has trained investors to overweight the importance of capital advantage. Particularly in deep tech, there’s a minimum amount of money needed to get to the next lamp post. Adding tens or hundreds of millions on top of this is a marginal benefit, and is generally not enough to offset more fundamental dynamics. I’m reflecting on this as I think back to some of the early partners I wanted to get on board and could not because of this capital advantage fear. I was a young upstart out of nowhere with very well funded competition. But in the last two years, the Valar team has made insane progress on 1/10th the capital we were told it would take. Now, because of that, we’re getting to a place where capital is easy to access too. Pretty soon we will have that advantage as well, as well as all the others. (I still don’t think it will be the most important). I think I feel compelled to write this out because it feels important to the soul of what makes the American tech ecosystem so great to course correct away from this. The argument can be made very selfishly: Valar will be a fund returner for those early believers, and there are others like it just getting started. But more fundamentally, the whole *idea* of tech investing is to find the Davids who are building slings. The fact that the Goliaths are more capitalized is what makes them juicy targets. VCs are beginning to sound more like bankers and less like pirates. This seems bad. We should figure out how to course correct from that. My favorite investor consistently reminds me: “There’s a lot of money in the world. You can have as much money as you want. Is that actually what’s blocking you right now?” Usually it’s not.

@andrewmccalip @redbullfuturist I think the appropriate answer is that powder bed isn’t scalable manufacturing. There are certainly levels to the competency of additive, and massive strides to be made this decade and the following.

@redbullfuturist We don't, because it's not a serious form of manufacturing.

@ranman @andrewmccalip @redbullfuturist Super interested in your background considering having a background at both PyTorch and SpaceX. Have anytime to chat?

@andrewmccalip @redbullfuturist We used additive metal manufacturing quite a bit at SpaceX. It's not crazy scalable but it's certainly useful.