Harrison Kinsley

testing robot policies has never been so much fun

@AndrewCurran_ Mythos didn't change anything. Mythos marketing doomhype did, maybe. IDK why we keep falling for this schtick. This is all purely because OSSAI is dropping models every week closer and closer to the flagships and even surpassing them in certain domains.

The Trump administration has informed Anthropic, Google and OpenAI that they are discussing the creation of new AI oversight procedures that would potentially require new AI models to pass a safety review before being cleared for release. Mythos has changed things.

@SanDiegoKnight he doesnt need to answer IP related questions. But it would be nice to give demos to someone who knows to ask why a robot begins turning before it was verbally commanded to do so...etc.

@Sentdex Genuine question: wouldn’t that possibly reveal too much about the tech and IP? This seems to be a race to first to market. Why give any serious insight or clues to the tech?

when will this man allow someone with deep RL, humanoid, && robotics industry knowledge to interview him

@BartronPolygon @LACYF50 The Dwarkesh video you linked is an example of the difference. Anyone who goes to a Dwarkesh interview can expect a shrewd and critical interview. Musk is at least open to this, even if you don't like his answers. It would be shocking to me if Adcock did a Dwarkesh interview.

@LACYF50 @Sentdex I don’t see this as qualitatively different than Brett’s videos. Both of these guys understand a great deal of what goes into the products their respective companies make and can talk about production and design challenges, cite specs and performance numbers, etc.

@BartronPolygon you may not like his responses, or find them satisfactory. but he gives responses and opens the door to scrutiny.

@Sentdex Really? Musk just rambles through very superficial technical points (“people don’t shoot lasers out of their eyes”, “you have to understand things at the limit”, “physics is really important”), stutters a bunch, and then tosses out ludicrous projections that end up being wrong.

@BartronPolygon musk will sit down & talk technical w/ anyone who asks him technical questions and is more than willing to defend his views. I understand why people still don't like him, but these 2 people are nowhere near the same. adcock is unwilling to even allow for the chance of scrutiny

@Sentdex Don’t really get the hate for this guy. Even Elon doesn’t do deep technical interviews because, like Brett, he’s a marketing guy.

The only time an interview has ever been given to anyone with industry knowledge was with @GoingBallistic5 in 2024 and the plan for actual AI in the robot was a mere concept of a plan with partnership thru OpenAI and MSFT. youtube.com/watch?v=Cv1TD2…

@zshobbs 2.7 will run on 2 comfortably

@Sentdex Do yu need 3 for m2.7 or will it run on 2?

While I also have a bunch of GPUs, I really enjoy the GB10s I have, which are the same as Sparks. Very low power usage, make no noise, generate very little heat, take up little space. Always on and available. 50 t/s with plenty of very solid MoEs.

@ZenMagnets @SullyOmarr google trends don't lie, but liars use google trends

@SullyOmarr And yet still indisputably in the lead

Unsurprisingly search for openclaw is basically dead

@mohits_tweets I thought this pic was a well known meme on tpot and I don't think OP was misleading at all lol. plus everyone knows my server rack is in my kitchen.

@Sentdex that's definitely not you

lightning knocked out a couple of my networking switches, found myself standing at my rack hands on hips trying to trace all the cabling and solve the puzzle of which switches were still alive and realized i was literally doing the meme

@kevin_zakka Very cool to see. I wonder if it would transfer

I trained a grasping policy on objects of various shapes and sizes using the new mjlab feature and this really cool pivot grasp strategy emerges for large flat objects that are larger than the gripper aperture. So cool and beautiful to see!

@KinvertOG got any pics u can share of the barn build btw? would love to see em

@Sentdex Yeah I can find 5090s fairly reliably around $3,500 right now. I got mine under $2,400 each. Was a bit late on the ram and nvme but did ok. Got enough to build another 5090 box if I need it. But lately I've been busy with a barn build and haven't been coding much.

i just want to open fbm and see someone selling like 8 really cheap rtx pro 6000s in rural alabama. is this too much to ask

@KinvertOG at least the market is calming down and coming down a bit lately

@Sentdex have Claude check every minute for it. They probably pop up twice a week and someone else gets them first. Like MSRP GPUs at MicroCenter.

@andrewc_actual @tomcocobrico almost certainly true, which is why I wish they were just explicitly forbidding the usecases they didn't want. At this exact moment in time though, i would still like to use their model. it's probably the best model at this time that's reasonable to run local in my eyes.

@Sentdex @tomcocobrico In a couple months it will probably be replaced with something else anyways 🤣

@olipdsanders Spark/GB10s themselves are whole ass computers that also serve many purposes besides LLMs too though.

@Sentdex With the RTX 6000 Pro you do need a whole computer around it but it serves many purposes besides LLMs.

@tomcocobrico @andrewc_actual They have put that in text on twitter yes, but if you read the license, this is prohibited and requires written approval. I was hoping they would stipulate clearly in the license you could use it commercially as a personal coding agent...etc, but nope not yet.

@Sentdex @andrewc_actual Did they not say already you can use privately hosted one for work outputs? unfortunately only have a single spark still and only running 2.7 int4 which i pruned with REAP. But it works

@ns123abc the moment the only remaining claims of the case were equitable, this was pretty much all set in stone. maybe we'll be surprised, but I doubt it.

🚨 OpenAI's lawyer in opening argument just showed jurors a 2015 Musk email: "Probably better to have a standard C corp with a parallel nonprofit" He's arguing Musk himself wanted OpenAI to go for-profit as early as 2015. What the court room reporters won’t tell you: The judge already addressed those exact emails in the January 2026 ruling, her own words: “Musk insisted throughout those discussions that OpenAI remain a ‘ philanthropic endeavor.” Musk floating structures in 2015 is not the same as Musk approving the conversion in 2025. He ended the debate in September 2017. Altman and Brockman told him they agreed. OpenAI’s lawyer is showing one slide of an eight-slide sequence. Judge Gonzalez Rogers saw all of this and let it go to trial anyway.

@cixliv @REK dude that's epic!

We just signed a lease for the first humanoid store in the USA! The @REK Shop. Coming to San Francisco Nob Hill in ~2 months. Shop during the day, robot fights at night.

@andrewc_actual 100% agree. It's not just an RTX Pro 6000. It's an entire computer that needs to also be built around it too. Very much in love with M2.7 as well. What a beast of a model. Trying to secure a license atm to use it with work

@Sentdex I like my sparks. 2x spark cluster runs minimax m2.7 awq at ~ 40 t/s. Alternative to that would be a dual pro 6000 rig - which is in a entirely different cost/noise/power range