Keaton Hare (@keaton_hare) header
Keaton Hare (@keaton_hare) profile picture

aerospace

Joined March 2020
Photo by Keaton Hare (@keaton_hare)Photo by Keaton Hare (@keaton_hare)Photo by Keaton Hare (@keaton_hare)Photo by Keaton Hare (@keaton_hare)Photo by Keaton Hare (@keaton_hare)Photo by Keaton Hare (@keaton_hare)Photo by Keaton Hare (@keaton_hare)Photo by Keaton Hare (@keaton_hare)Photo by Keaton Hare (@keaton_hare)Photo by Keaton Hare (@keaton_hare)
Finally got around to editing together flight 12 footage - you can just catch my foot slipping as I switched to over the shoulder shooting…
5
5
66
8,710
Keaton Hare retweeted
*asking if someone is ingroup in san francisco* is he a claude fearing man?
11
56
2
972
14,588
Keaton Hare retweeted
normal person reading Demis Hassabis' biography: you can just accept two postdoc offers at the same time ryan: you can just win the nobel
Hey bio twitter, I was sick in bed this weekend but luckily my clankers were clanking This time they found the first known case of genome-wide frameshifting in bacteria! A whole order of gut bacteria skips a letter of its genes at CGG. How they found it: My agents aligned one genome from every bacterial and archaeal family in GTDB to known proteins, this time allowing for frameshifts. Every genome has a few apparent frameshifts (assembly errors, dying genes), spread thinly over all codons. In one TANB77 (uncultured Clostridia) genome they piled up at CGG. So we tested every TANB77 genus, finding that: - 54 of 64 genera show the skip, and all 54 have lost the CGG tRNA - Related proteins have glycine at 1,237 of 1,458 skip sites, vs 6% at other frameshifts - 59% of all +1 frameshifts in TANB77 fall exactly at CGG (1,026 of 1,744); the rest are the usual background noise, scattered thinly over all the other codons. - Their phages skip CGG too: 58% of phage CGGs sit where a skip is needed, vs 2% in phages of other Clostridia - It's not a sequencing error or RNA editing: DNA reads (140 of 140) and gut RNA (305 of 305 read pairs) keep the C Closest prior work: Genome-wide frameshifting is known in the ciliate Euplotes, and glass-sponge mitochondria skip a U the same way, but at only 2 sites.
2
4
1
204
33,907
Keaton Hare retweeted
Important peer-reviewed research result:
26
89
14
1,303
57,333
Keaton Hare retweeted
I have a major result to share. I’ve found conclusive forensic evidence that the top 4 Chinese AI labs—DeepSeek, Moonshot, MiniMax and Ant Group—have built their newest models on NVIDIA’s most advanced Blackwell chips, which are barred from export to China. This sounds insane, but I discovered that almost every AI model carries a hidden signature of the GPU that trained it. When a model is first set up, the starting values of its weights are laid out in a pattern that depends on the exact GPU. Each combination of GPU and seed number produces a unique pattern—a sort of fingerprint of the hardware used to create the model. Traces of that pattern survive training into the published model. Again, this evidence isn’t based on any insider information or leaking. It’s based on a forensic analysis of the model weights released by the companies themselves. I searched more than six trillion combinations of chip layout and seed number and matched them to open-weights models. In doing so, I found conclusive evidence that six Chinese AI models—including DeepSeek-V4-Pro, Kimi K3, and MiniMax-M3—were initialized on NVIDIA Blackwell systems. Four of the models match Grace Blackwell, the GB200 and GB300 parts in NVIDIA’s flagship liquid-cooled rack-scale systems. I then verified these results on real hardware. Running unmodified PyTorch with the seeds I’d recovered, a Blackwell B300 reproduced DeepSeek-V4-Flash’s untouched starting weights bit for bit, and a rented GB200 reproduced Kimi K3’s to a correlation of 0.999995. A Hopper H100 reproduced none of DeepSeek-V4-Flash’s, and changing the seed by one broke the match. (DeepSeek’s seed is 953049583 by the way.) The method also was verified to match 24 known open-weights models to the hardware that developers said they used. It is exceedingly unlikely that these results are a coincidence. Across six trillion combinations, the best accidental match should reach only about 8 sigma (i.e., eight standard deviations above what random chance would produce). The real matches range from 49 to 1,385 sigma, and even the weakest single test hits 20. These were data-center-scale training runs. With some analysis of release schedules and model details, I estimate that the largest, Kimi K3 and DeepSeek-V4-Pro, used 2,000 to 3,000 Blackwell GPUs each. Clusters of this size cost on the order of $150 to $200 million each to buy, and across these labs alone would add up to more than half a billion dollars of export-controlled GPUs, not counting experimental compute or otherwise. It’s quite likely that we see multi-billion dollar expenditure on export-controlled GPUs. I am highly confident in this result, and it has been independently verified by senior researchers at OpenAI and Anthropic. You can read the full report at unemploymentlabs.com/checkpo…. There is much more I’ve identified about these models through other forms of checkpoint forensics like this, and will have more to share soon.
1,817
6,412
2,191
41,171
6,651,980
A thread worth reading
Oh I suppose I should tell the story of one of the most remarkable training requirement stories I’ve ever heard of. After the Vietnam war, all four branches of the US military that were involved went through a period of soul searching. This period led to a number of reforms.
2
292
Keaton Hare retweeted
Oh I suppose I should tell the story of one of the most remarkable training requirement stories I’ve ever heard of. After the Vietnam war, all four branches of the US military that were involved went through a period of soul searching. This period led to a number of reforms.
1
8
3
150
21,888
Zipline is going to be much bigger than this one day
Today, Zipline raised $1.5b at a $21.5b valuation. When we started Zipline, flying autonomously in the US was illegal. Which made it really hard to convince investors to give us money! But we were too naive to think that meant we shouldn’t try. We started the company on a cow farm. We worked out of run-down construction trailers so that we could keep engineering, manufacturing, and flight test all unified and in one place. Hiring people from normal companies like Google was hard because we didn’t have a cafe or laundry service. But enough risk takers and rebels wanted to go on an insane adventure with us. They became the early team. We decided to launch in smaller countries and focus on life-saving use cases. We hired incredible entrepreneurs and operators in Rwanda, Ghana, Cote D’Ivoire, and other countries. For eight years, Zipline expanded our service across Africa, serving thousands of hospitals and health facilities and reducing maternal mortality by 51%. Today, the network saves over 14k lives every year and is cash flow positive. Last year, Zipline launched at scale in the US and the service exploded. We grew flight volumes about 100x over the last 18 months, and now make more than a million deliveries a year per metro. We’re now expanding with our partners to millions of deliveries per day. Normal families love having access to anything they need in 10-20 minutes, no matter where they live. Spend more time with loved ones and less time worrying about errands. Zipline builds drone delivery that is silent and safe. No one wants to hear an extremely loud “swarm of bees” delivering to their backyard or their neighbor’s. Automated delivery needs to be insanely quiet, which guided the overall architecture of our technology delivering from 330 ft in the air. We are 6 times quieter than the closest competitor during delivery. Zipline also now has over 150m commercial autonomous miles (10x the rest of the industry combined) and no safety incidents. Neighborhoods should demand silent and safe delivery! Today, Zipline is announcing two things. First, we just signed the largest autonomous vehicle contract in human history to expand automated delivery nationally with Walmart, our first and biggest partner in the US. We’re starting by scaling to 300 supercenters throughout the country. Second, we’ve raised $1.5b at a $21.5b valuation from existing and new investors, led by Paradigm, WCM, and Layer. Zipline was a hard company to believe in for the first ten years of our history. Investors didn’t think the company was “investable”. Most people didn’t even think Zipline was an economic venture. Thanks to those who have rooted for us along the way. To all of the tens of millions of customers and families who rely on Zipline day in and day out: thank you for helping us build the future. Zipline has grown more than @ryanzip, @keenanwyrobek, I, and all others at the company could have imagined. There is a big opportunity to build a new layer of infrastructure for humanity over the coming years. We want to approximate teleportation. We can build a logistics network that connects and serves all people on earth. We can free human flourishing from the choking constraint of traffic and roads. There are no roads where we’re going. 🚀🌎
1
7
468
Keaton Hare retweeted
🎉 Hosting a Hardtech Bonfire Night in SF Bay Area, Saturday, 11/14 **At original TESLA HQ where roadster was unveiled** 👥 We’re bringing in 200+ of the best founders, engineers, VCs and patriots working in the hardtech space to talk robotics, defense, nuclear, drones and everything in atoms. Whether you are building a company, an engineer looking to get hired or looking for a company to invest in, we want you here! Grab a beer and meet the most interesting people in the Bay over a bonfire 🔥 🤖Bring your robots, drones, flying cars… whatever you’re working on! We want to see them. There will be some surprises! 🎤 Top 10 best teams get to pitch their idea to over 200+ people 📍Location: OG TESLA HQ - 1056 Commercial St, San Carlos, CA 94070 🚊 20 min train ride from Soma 🍻 All food and drinks covered 💻 Fill out Luma link, comment and DM if interested! We’re going to fill this out! **Looking for sponsors! DM if interested!** luma.com/ak5cch9g
40
14
10
124
31,672
please form an LA based 24 hours of lemons Racing team with Lucas We will duke it out SF va LA on October 2nd, 2027, at Buttonwillow Raceway.
Who’s going to build a gundo 24 hrs of lemons car with me?
1
3
16
821
Keaton Hare retweeted
So I recently badgered my cinematographer @steveyedlin into making my ideal iPhone still camera app. It shoots RAW, bypasses the iPhone's look processing and instead uses the same image processing we use on our movies. apps.apple.com/app/apple-sto…
813
2,851
2,530
57,003
14,150,469
Broke all the rules for this one. Absolutely STOKED to make this partnership official. It all started when I went to Jim for help with an anodizing line that I was considering buying. He told me I was nuts, I soon realized I was nuts, and a friendship was born. Investing in Jim might have been one of the easiest decisions I’ve ever made. He’s the EXACT type of founder and business I want to be a part of. What he has built is truly incredible. And the best part is that Jim and SCS are just getting started. @CRV is honored to be on the ride. LFG!!!!🔥🔥🔥
Okay it’s happening: - Added 160k square feet in Reno for our automated high volume lines and parts up to 5’x10’. Operational now. - Added 170k square feet in Pittston PA for 1-day shipping (or faster) to the Northeast US. Landing machines and growing the team now. - Southeast US is next (flying out next week to negotiate a lease). 3 more leases in 3 more states to sign after that. Expect massive lead time reductions across our factory network. - Finalizing our first acquisition and working on the second and third. These are services you’ve been asking for. More coming. We’ll be able to say yes to more of your BOM. - Welding is ramping nicely. More assembly support soon. - More materials and capabilities for CNC machining in development now, including much smarter pricing. Increased capacity will also help us cut lead time drastically. - Added 450 people so far this year. Hopefully double that in 2027. - Candy and merch budgets have been tripled. …And we’re just getting started. That’s a big part of why we raised again: to build more, ship faster, and make manufacturing easier and more accessible. I’m doing my best for ya. I love you guys. Big thanks to @willreed at @sparkcapital , @caitlinbolnick1 at @CRV, @marksugarman at MHS, @andrew__reed at @sequoia and @matthuang at @paradigm for supporting our Series B (and the extra rocket fuel) LFG!!!!! 🇺🇸
23
3
3
139
14,971
Keaton Hare retweeted
I spent a year working the apps for DoorDash, Uber, Amazon and more; saw disturbing things; brought people single cookies; probably delivered drugs; smashed up my nose; met people who made insane journeys to get you pad thai; wondered what we’re doing here simonandschuster.com/books/D…
26
166
18
2,769
609,996
Keaton Hare retweeted
I was an 18-year-old girl in biotech who had admired Elizabeth Holmes. The day after she was convicted, I wrote this Some Thoughts on Theranos, 01.04.22
8
12
1
197
18,093
Keaton Hare retweeted
~ $0 raised, btw. Cash flow positive in month three, no outside investment except $180k from my Dad and a HELOC on my house to buy our first laser.
OSH Cut is accelerating. We're on track to double revenue this year, and we're adding capacity as fast as we can to keep lead times short!
47
11
6
554
17,039
Birds of San Francisco
1
21
547
Keaton Hare retweeted
have some lore to share: the first vibecoding company was not cursor or cognition or anthropic -- it was Season 6 of Silicon Valley > be me, techbro founder hired as a consultant to the show
21
40
10
950
55,986
Keaton Hare retweeted
43
244
13
2,716
45,338