Luke (@groccy1) header
Luke (@groccy1) profile picture

🇺🇸🇯🇵🇹🇼 building the AI of tomorrow from first principles, prev PhD @MIT @UIUC EECS #TSLA $TSLA $SPCX 🚀 🚀 🚀 🌕. YT: Plaid Luke

San Jose, CA
Joined September 2019
Photo by Luke (@groccy1)Photo by Luke (@groccy1)Photo by Luke (@groccy1)Photo by Luke (@groccy1)Photo by Luke (@groccy1)Photo by Luke (@groccy1)Photo by Luke (@groccy1)Photo by Luke (@groccy1)Photo by Luke (@groccy1)Photo by Luke (@groccy1)
Loves Elon's nerd mode!
Replying to @BLKMDL3 @wholemars
In case anyone is curious: We’ve had a ~pixel-perfect world model since last year, which is critical for synthetic data generation. The (much) smaller version of that in the car is a conditional mean-flow model, based on the recent sequence of Tesla Universal Vision Encoder features, which are the same ones that are plugged into the 3B encoder model that produces the driving actions. We fork off from these UVE features, then run a separate ~1B diffusion model. This takes in the previous reality-view generation with added noise and then denoises it to the current visualization. It then produces a small video chunk (8 images at 36 fps, at the resolution of the UI screen). This work is done by 1/2 of a TRIP engine every UVE frame at 18 hz and is spread over 4 cycles. Currently, the latency is a bit high due to compute constraints, but we have ways to interleave newer UVE features along the diffusion transformer layers to cut roughly 35% of the latency.
10
Luke retweeted
v15 is 25x safer than v14 across the fleet. That’s hard to notice on any individual drive because most drives are already very safe. The difference only becomes clear when you aggregate millions of miles. All credit goes to the incredibly talented and dedicated @TeslaSI engineers who made this possible. It’s an honor to work alongside them.
FSD v15 now starting to roll out – upgrading every part of the end-to-end driving network Safety 25x better handling of safety critical cases Tesla Reality View  Adds a live visualization of the drive that lets you see & control what FSD is planning to do Navigation for start & end of the trip now learns end-to-end from fleet route preferences. Later v15 releases will expand this to the entire route. You also should notice better detection of signs, animals, debris, potholes and standing water, plus more natural speed handling in school zones and inclement weather. Lastly, FSD can now slow for speed cameras & Bike Rack Mode lets you enable FSD w/ bike racks or similar hitch accessories
232
616
77
6,849
669,679
The power of Tesla SI
Replying to @lawsonfulton
And we only started working on it a few months ago 😂
11
Right now, the media is full of stories that paint Jensen Huang in a bad light. I have actually met Jensen, and I want to tell you a story. (1/2) My interview with Jensen hasn't been published yet, which is probably understandable given the current situation. More on that when the time is right. But I had the chance to meet Jensen Huang in person several times while traveling through Europe with NVIDIA, and I want to tell you about it, especially about one experience that has stayed with me. Some quick background: NVIDIA invited me to join Jensen's European tour last month, together with Bloomberg and Steve Nouri. We were there for every stop: Paris, London, Stockholm, Berlin. I was on site for every one of his public speeches. A lot is being said right now about Jensen making statements that many find contemptible, while others take them with humor. And to be fair, Jensen does have a slightly edgy humor that rubs some people the wrong way. He speaks plainly, says what he thinks and makes jokes that can offend, but he always makes fun of himself too. And I, at least, can say one thing very clearly: never once did I feel that any of his remarks were meant as an attack, as belittling or as humiliating. It was always situational comedy. If Jensen hadn't been CEO of NVIDIA, he would have made an outstanding stand-up comedian, always ready with a quip. But as we all know, that's a matter of taste, and saying that I always found it amusing won't calm anyone down. So here are two stories that show Jensen really has a good heart, and that the campaign against him distorts his image. This is my personal experience: NVIDIA gave me the opportunity to do a 30-minute interview with Jensen in Stockholm. As you can imagine, I was incredibly nervous. Honestly, I was even scared. Me, a small "content creator," getting to meet Jensen in person and talk to him. He makes time even for little people like me. I met him and his staff at a café in Stockholm. We went inside and, as was bound to happen, my camera gear fell out of my backpack onto the floor. My stress level: through the roof. But Jensen stayed calm and relaxed, told me everything was fine and that I didn't need to stress, and put his hand on my shoulder. Not one snide word, no making fun of me. Quite the opposite. He knew I was genuinely stressed and helped me relax. Then we started the interview. And while we were talking, people kept coming into the café where we were recording. It was a café open to the public. And as you'd expect, the same thing happened again and again: while Jensen and I were talking, guests recognized him and more or less burst into our interview. And Jensen took time for every single one of them. Not one bad word, no irritation, no stupid remark. He gave every guest his time, took photos, had conversations. In the middle of the interview. I was surprised and genuinely moved. No arrogance on his part, just friendly humanity. And that's how an interview scheduled for 30 minutes turned into 60. Simply because Jensen and I had such a good conversation and he kindly gave all those guests extra time on top. But the most impressive part actually came afterwards. We had been eating a Swedish pastry, it's called “fika” (at least that’s how we called it). And Jensen loved it. After we wrapped up the interview, Jensen went to the trunk of his car, grabbed some cash, went back into the café and bought “fika” for everyone. For himself, his security, his staff, for everyone. No camera recorded it. Nothing done purely for PR. Nothing meant to make him "look better." Just a human gesture, because he liked the pastry and wanted everyone to enjoy it too. If I hadn't told this story today, it would never have been told. And to me it shows that Jensen is a genuinely good person. I gain nothing from writing this. Jensen is no longer CEO. And nobody suggested I write this story down.
227
490
38
4,439
209,422
Luke retweeted
This is real-time AI video generation (no gaming engine!) using just half of one AI4 TRIP accelerator. Let that sink in …
The crazier thing is that the right side uses less compute than what the left used to take. Extreme compute efficiency optimization.
1,319
1,856
380
20,882
2,908,339
The crazier thing is that the right side uses less compute than what the left used to take. Extreme compute efficiency optimization.
Look how far we’ve come
843
1,396
297
16,275
1,726,262
I've never known the tech world to be more unanimous about anything than they are about Jensen Huang. Everyone in tech thinks the board made a huge mistake. Except the board. But they are not in tech.
1,423
1,907
397
26,922
1,712,953
TLDR of this awesome read: 1> Most harnesses keep appending the whole history into one prompt, so the model has to cope with contexts it never trained on 2> Harnesses like RLMs instead keep the context outside the prompt, write code over it, and hand smaller pieces to sub-calls. Each LLM call stays "locally in-distribution". Trained only on short tasks, this generalized to tasks 2–64x longer and transferred to new domains with similar structure. 3> The cool trick is surrogate harness training. They train the orchestrator on cheap synthetic environments with fake tool calls and fake sub-agents instead of expensive real rollouts 4> The resulting RLM-27B beats base Qwen3.8-27B by 26.9% on average across 8 hard long-context benchmarks (58.5 vs 46.1), while using about 2x fewer tokens take: models won't eat the harness, we just haven't explored what harnesses can do yet
Harnesses used to be a crutch for weak LLMs. In 2026 and onward, with stronger models, harnesses now offer a mechanism for composing an LLM’s existing and new capabilities to generalize well beyond their training environments. In our new paper, LM Harnesses are Compositional Generalizers, we explore how various harness abstractions enable LM systems to generalize over task length, task strategies, recursive task depth, and more through composing existing LLM capabilities. These abstractions provide a means for more data-efficient training, especially over long tasks, by keeping trajectories locally in-distribution (LID) — i.e., shaping all individual LLM calls to be in-distribution. For tasks that take a very long time to roll out, this implies that we can even perform surrogate harness training, meaning we train an individual LLM call instead of doing the very long harness rollout, using purely synthetic tasks, with fake tool calls + sub-agents! A common idea floating around is that LLMs will absorb harnesses and make them obsolete. We disagree, and argue that we haven’t sufficiently explored what harnesses can provide. The full paper, logs, codebase, and data are available: alphaxiv.org/abs/2611.harnes… w/ @SOURADIPCHAKR18 @lateinteraction
7
9
2
231
14,425
Will stronger LLMs make harnesses obsolete? New MIT CSAIL paper argues the opposite: Harnesses compose model capabilities to generalize across longer, deeper, & more strategic tasks, while enabling more data-efficient training: tinyurl.com/2fwk4uap
15
42
5
275
14,512
Harnesses used to be a crutch for weak LLMs. In 2026 and onward, with stronger models, harnesses now offer a mechanism for composing an LLM’s existing and new capabilities to generalize well beyond their training environments. In our new paper, LM Harnesses are Compositional Generalizers, we explore how various harness abstractions enable LM systems to generalize over task length, task strategies, recursive task depth, and more through composing existing LLM capabilities. These abstractions provide a means for more data-efficient training, especially over long tasks, by keeping trajectories locally in-distribution (LID) — i.e., shaping all individual LLM calls to be in-distribution. For tasks that take a very long time to roll out, this implies that we can even perform surrogate harness training, meaning we train an individual LLM call instead of doing the very long harness rollout, using purely synthetic tasks, with fake tool calls + sub-agents! A common idea floating around is that LLMs will absorb harnesses and make them obsolete. We disagree, and argue that we haven’t sufficiently explored what harnesses can provide. The full paper, logs, codebase, and data are available: alphaxiv.org/abs/2611.harnes… w/ @SOURADIPCHAKR18 @lateinteraction
74
219
53
1,559
193,941
Luke retweeted
Went on Danish national TV yesterday to advocate for the sensible return of Jensen. It's a common misconception in Europe to worry about "tech power", as though Jensen stole it from someone else. No, he damn well built it by putting in 33 years of service on behalf of NVIDIA.
243
189
21
3,994
179,714
Oh my, FSD in Italy! 🤯🚀🚀🚀
Mamma, butta la pasta Now approved in Italy 🇮🇹 Rollout begins soon
20
Absolutely the future I look forward to. I would want my company to do the same.
Replying to @YejinChoinka
@YejinChoinka and I joined @NYSE Live with @KristenScholer to talk about @noeri_ai, the company we co-founded and announced on Monday. At the core was a simple question: what is Noeri? Today, companies face a trade-off: the best model is expensive – the cheaper one gives worse results. Either way, they rent their AI from a handful of labs, and the bill grows with every token. We’re building self-improving AI that you control and own. It specializes in your company’s own work and keeps getting better at it, so results go up while cost comes down. And the agents you run today learn new tasks. Because you own it, your AI stack becomes an asset on your balance sheet, not rent paid to model providers.
5
We got your back @JensenHuang !!
TRAVIS KALANICK: "DON'T LET NVIDIA GO TO THE SUITS, JENSEN" (FULL INTERVIEW) 1:30 - Jensen Huang's ouster was an op 4:20 - Playbook for fighting back 9:15 - "Jensen Huang is a good person who made mistakes" 10:40 - Don't let NVIDIA go to the suits 12:00 - NVIDIA's next CEO will try to keep Jensen out 13:25 - Travis is betting on Jensen's return 17:45 - NVIDIA employees should walk out in support of Jensen 19:50 - Opposition to data centers is a red herring
116
304
16
4,998
135,888
I’m just going to leave this here in contrast to the mainstream media’s current work on @JensenHuang. He was voted #1 ceo to work for in the world this year, with CEO approval rating of 99%. Employees believe in his vision, in America leading in relentless innovation, and in an open AI ecosystem.
219
746
58
5,635
257,027
Ok.. I get @JensenHuang "Deer in the Headlight" now... I think everyone owes him an apology 👀
Nvidia interim CEO Colette Kress says the board has determined it will not reinstate Jensen Huang “Jensen resigned from his position with NVIDIA, and the board accepted his resignation.” “The board has determined that they will not reinstate Jensen. And that is our statement and what the board has decided to do.” “We wanna think about right now growing our business. We wanna grow in terms of everything that we need to do with our employees, our customers, our shareholders, and that is where we are gonna focus moving forward.”
39
45
2
672
20,986
Luke retweeted
Many of us who have worked corporate jobs and have seen such cold & soulless management know how suffocating these ppl make an organization. No way you’re inspiring anyone or retaining talent with that tone, look and callous attitude. What do you think Nvidia talent feels like right now watching this? They see that their inspirational founder is thrown under the bus and his 33-year effort means nothing to this management. You think they will feel safe and inspired? She could have said this in so many better ways.
Nvidia interim CEO Colette Kress says the board has determined it will not reinstate Jensen Huang “Jensen resigned from his position with NVIDIA, and the board accepted his resignation.” “The board has determined that they will not reinstate Jensen. And that is our statement and what the board has decided to do.” “We wanna think about right now growing our business. We wanna grow in terms of everything that we need to do with our employees, our customers, our shareholders, and that is where we are gonna focus moving forward.”
699
870
77
10,150
538,393
Only three people to remove from the board? Easy work 15% shareholder vote coming
216
552
45
9,517
379,830
A special week for Simile: our first company-wide offsite! Truly grateful to work with such an inspirational team every day. If you’re interested in joining one of the most talent-dense teams in tech, reach out. We’re hiring!
Simile’s first 100 at our first offsite. For a team working to understand human behavior, it’s special to spend time getting to know the humans behind it all. What started as a small team with an ambitious vision has grown into an incredible group from OpenAI, Google DeepMind, Stanford Research, Harvey, Figma, Glean, and beyond. To our first 100, we’re so glad it’s you.
5
3
94
5,771