Haydon Ryan

Builder, Lover of all things automation - software, hardware. Make the machines do the work so we can get more out of life! Opinions are my own

[Haydon Ryan @haydonryan]

@cyris_kissane Not using NVLink on the H200s?

[Haydon Ryan @haydonryan]

@haydendevs Me too!

[Haydon Ryan @haydonryan]

@droidbuilds Flux .dev , Stable Diffusion, Openwebui, vllm, tensorflow.

Depends what you define as a "tool"

[Haydon Ryan @haydonryan]

@Bhavani_00007 OpenAI supports opensource. Anthropic are just in it for themselves.

[Haydon Ryan @haydonryan]

@dreamsofcode_io Time to rename yourself to Dreams of Tokens ;)

[Haydon Ryan @haydonryan]

@gospaceport IMO it's still Tiering. I've found bunch of situations where I need more. It's probably 80/20. Especially when it comes to planning or the "other" side of dev - refactor, optimization etc. Some very heavy stuff kimi won't do and only sota can (<1%). Q2.6 -> K2.6 -> GPT5.5.

[Haydon Ryan @haydonryan]

@atmoio I swear Uber (and all these other companies) needs to just buy some local hardware.

[Haydon Ryan @haydonryan]

@AllieJade1 Can't wait for auto pilot for everyone. Having driven in Germany, this would drive me nuts!

[Haydon Ryan @haydonryan]

@chrishume_ @CaitlinCloud9 Nah it’s mild today in Austin due to crazy rain. Not looking forward to it heating up though

[Haydon Ryan @haydonryan]

@snwy_me Next we’ll see method unlocked my seat heaters in my BMW. I hadn’t actually contemplated that metho is gonna find a lot of vulnerabilities in existing hardware.. Very interesting post thank you!

[Haydon Ryan @haydonryan]

@QingQ77 Love this apartment setup

[Haydon Ryan @haydonryan]

@FanlessTech I feel like this should have been flipped over the other way - no looping cables, and cpu heatpipes venting to clear air.

[Haydon Ryan @haydonryan]

@neogoose_btw Time to try an autoresearch loop on your usecase to see if you can improve performance.

[Haydon Ryan @haydonryan]

@RoboIntellect @NVIDIAAI @openclaw @NousResearch @LangChain @huggingface If you were ok running slower then you could get away with less GPU as long as you had enough system ram.

[Haydon Ryan @haydonryan]

@RoboIntellect @NVIDIAAI @openclaw @NousResearch @LangChain @huggingface 550B at nvfp4 is ~225gig of weights. MOE means if you have fast enough system ram you can swap in and out of vram but it would hurt performance. 4x rtx6000 pro would run it (384gb) without swapping ~ $50k USD.

[Haydon Ryan @haydonryan]

@david_nix @gospaceport @loktar00 I got mine in the 8xxx I now want another. I hope your Time Machine has room for two. If I’d known about pipeline parallelism then I prob would have bought two.

[Haydon Ryan @haydonryan]

@wagslane The old SEO approach with click through ads will not work though. I can see a model where you pay $1-3 a month for read only bot access, and maybe $50 or per api call for write (limit bot slop) in the big platforms. Companies who don't adopt this will be at risk of disruption.

[Haydon Ryan @haydonryan]

@wagslane Yes - I can totally see there being tension between what people want (via bots) and established companies. This will spurr competitors that are bot friendly. I would argue you can "sell" to bots but it's more when the person gives a "research" request to the bot. ...

[Haydon Ryan @haydonryan]

@theo This is not an intractable problem to solve.

[Haydon Ryan @haydonryan]

@theo Feels like a need for a new OSS compeditor. I know mattermost is reasonable. Esp with agentic engineering - something user friendly would be great. Teams only benefit is if you have a teams meeting and the chat stays in the teams app, vs zoom where it disapears.