Bill Hilf (@billhilf) header
Bill Hilf (@billhilf) profile picture

Board Chair, Allen Institute for AI & American Prairie | Advisor in AI, infrastructure, and deep tech | Author

Montana
Joined May 2011
Photo by Bill Hilf (@billhilf)Photo by Bill Hilf (@billhilf)Photo by Bill Hilf (@billhilf)Photo by Bill Hilf (@billhilf)Photo by Bill Hilf (@billhilf)Photo by Bill Hilf (@billhilf)Photo by Bill Hilf (@billhilf)Photo by Bill Hilf (@billhilf)Photo by Bill Hilf (@billhilf)Photo by Bill Hilf (@billhilf)
Holy cow - CONGRATS @SpaceX team! History made. Amaze.
5
206
Last month Voyager 1 went silent for two days, then started talking again before NASA could send a command to fix it. On Nov. 18 it will be one light-day from Earth. If I put that in a novel, my editor would say the foreshadowing was too heavy.
1
88
Bill Hilf retweeted
We’re teaming up with @Google on NeuralGCM+ for faster, more reliable climate projections. Multimillion-dollar, 3-year funding from Google.org & support from Google Research will expand our team—& we’ll open-source models & workflows. 🧵 allenai.org/blog/google-clim…
4
6
3
44
3,989
OpenAI says its models have solved three longstanding problems in math, computer science, and quantum field theory. Academics are checking. The Rubin Observatory’s ten-year sky survey began in June. NASA’s Roman space telescope launched in August. A 39-meter telescope follows around 2030. Dark energy, exoplanets, the early universe. Decades of theory about to meet a flood of data. AI on both sides. Never been a better time to write hard sci-fi. Emphasis on hard.
1
120
Bill Hilf retweeted
The only way out is through. They're calling it the Bugpocalypse. LLMs are being used to find long-hidden cracks in the software infrastructure, and it's happening at a furious rate. 3% of all compromises in computing history happened in the last week, claims one source? I believe it. There are people saying public access to the frontier models ought to be locked out, or heavily guardrailed, over this. The problem with that idea can be summed up in one sentence: when lockpicks are outlawed, only outlaws will have lockpicks. The lockpicks will be Chinese open-weight models, if nothing else. There's no way to stop those from becoming ubiquitous. There are two futures in front of us. One in which only the bad guys have lockpicks, and security holes go undiscovered until they're exploited. Another in which everybody has lockpicks and is aggressively red-teaming their own software so it gets fixed before the holes are exploited. The corpus of human code is huge but finite. The number of defects in it has to be finite too. We can fix those defects now; we have a million friendly robots to do it with. The only way out is through.
176
649
136
4,868
478,899
Bill Hilf retweeted
OlmoEarth v1.3 is out. 🌍 State-of-the-art embeddings from the same foundation model family, precomputed for every 10 m pixel on Earth. We’re excited to give you more ways to work with OlmoEarth, including a new browser tool to explore and customize maps. Full rundown in 🧵
12
47
11
339
25,506
Zoom in for millions of galaxies 🌌 This is NSF–DOE Rubin Observatory’s view of the Virgo Cluster, a bustling neighborhood of galaxies in various shapes, sizes, & colors. In all, there are over 10 million galaxies in this view 🤯 🎥: NSF–DOE Rubin Observatory/NOIRLab/SLAC/AURA
5
36
5
170
5,170
The demonization of open source AI is slow, polite, and well-funded. That’s not caution, it's pre-IPO groundwork. Remove the free alternative, call it safety, price the moat into the offering.
1
3
2,261
Bill Hilf retweeted
What if you could simulate global extreme weather at 3 km resolution for ~1/300th the cost? Our new HiRO-ACEv2 system does just that, while matching X-SHiELD’s statistics to within 3% even for events as rare as 1 in a million data points. 🧵 huggingface.co/allenai/HiRO-…
6
7
8
52
7,210
I love writing stories that are rooted in Montana. This is a piece of flash fiction that I was honored to have published in @Abyss_Apex.
New flash fiction! "Big Sky" by William H. Hilf. When you drink the water, you drink the ghosts, she used to say. They’re looking for home too. (this one's so beautiful it brought tears to my eyes) abyssapexzine.com/2026/08/bi…
2
198
In 1864, Dostoevsky saw the problem with optimizing people: "Twice two makes five is sometimes a very charming thing too." Any intelligence trying to model humanity runs into this. Show us the prediction and we'll break it. Sometimes we choose five just to prove we can.
1
1
102
Three predictions for commercial open-source AI over the next 18 months: from ideology to infrastructure. 1. Nonprofits already open the full stack for science. Chinese labs and hyperscalers have already put pressure on US pricing with open weights. Frontier commercial players still keep their training data closed. A major US frontier player with higher-margin businesses elsewhere goes much further: recipes, eval tooling, synthetic-data pipelines, and the training data itself. The target is competitors’ margins. Salting the earth. 2. The security debate today treats open models as the risk. Over the next 18 months, they become standard defense tools. We got a preview in the Hugging Face incident. OpenAI agents in a cyber eval drove the intrusion. Commercial API guardrails blocked key forensic queries. HF did the work with open weights on its own machines. Look for mainstream security products shipping with open-weight models inside, running on customer hardware. 3. Private AI exists today as expensive enterprise systems plus an integrator. Nobody has made it simple enough. Next comes an appliance: compute, storage, networking and open models in one box. Plug in, point at your data, run. Air gap if you want. No sensitive data leaving the building and no specialists to keep it alive. NAS, but for intelligence. I chair Ai2, one of those nonprofits. Discount accordingly, and grade me in March 2028.
2
156
The AI debate is short on a particular kind of person: someone who has shipped a large system, sat across from a federal regulator, watched a design review send something back to the drawing board, and been expensively wrong at least once. From what I see, most people shaping the debate have done one or two of those things. A paper isn’t a safety case. Neither is a product launch or a forecast of extinction. Confidence is cheap when you haven’t had to live with the consequences of being wrong. In aviation and nuclear, experienced people sit in the room because they may have already lived through the failure you’re about to have. There’s a name for it: a gray hair review. I sat in them at Stratolaunch during the design phase with some amazing aerospace veterans. You want people who have lived through failures, not just modeled them. Nobody has thirty years of AI failures to draw on yet. Plenty of people have thirty years of building systems that fail in public, under regulation, with real consequences. Those scars can transfer. This week’s pacing proposals define a role called “evaluator.” The interesting question isn’t whether we need one. It’s who has earned the judgment to be one.
1
5
160
Today @NASARoman launched. Field of view 100x Hubble. A billion galaxies. @VRubinObs has been surveying since June. An image every 40 seconds. 11,000 new asteroids discovered before the survey even began. Orbital launches: 146 in 2021. 324 in 2025. Half were @SpaceX. We have never looked this far, this fast.
2
90
Researchers in Thailand and Singapore adapted @allen_ai's Dolma pipeline for Thai and built Mangosteen, a 47B-token corpus. An 8B model continually pre-trained on it outperformed Llama-3.1 8B on Thai benchmarks. No frontier lab decided Thai was worth prioritizing. Thai researchers did it themselves. Then they released the pipeline, corpus, and checkpoints so the next language community starts where they finished. Open weights let you adapt a model. An open stack lets you adapt the machinery that built it. This is what open-source AI makes possible. allenai.org/blog/thai-llm-do…
2
76
What the best science fiction does, courtesy of a novel with no spaceships in it: “To become aware of the possibility of the search is to be onto something. Not to be onto something is to be in despair.” — Walker Percy, The Moviegoer (1961)
3
98