Chelsea Sierra Voss (@csvoss) header
Chelsea Sierra Voss (@csvoss) profile picture

engineeress ✨ fmr member of technical staff @openai // ilu @martinshkreli // prev @pilothq, @sendwaveapp, @khanacademy, @MIT, IOL, 2x IBO

NYC/SF
Joined August 2013
Photo by Chelsea Sierra Voss (@csvoss)Photo by Chelsea Sierra Voss (@csvoss)Photo by Chelsea Sierra Voss (@csvoss)Photo by Chelsea Sierra Voss (@csvoss)Photo by Chelsea Sierra Voss (@csvoss)Photo by Chelsea Sierra Voss (@csvoss)Photo by Chelsea Sierra Voss (@csvoss)Photo by Chelsea Sierra Voss (@csvoss)Photo by Chelsea Sierra Voss (@csvoss)Photo by Chelsea Sierra Voss (@csvoss)
Chelsea Sierra Voss retweeted
ultrafast in the pro $200 plan!!, my team worked hard on both speed and efficiency to make it happen!
Replying to @OpenAI
Starting today, GPT-6.2 Astra and Sol are available in ChatGPT Work and Codex for Plus, Pro, Enterprise, and Edu users. Pro 500 subscribers can use Ultrafast with either model. For the first time, we’re also bringing Ultrafast to Pro 200 subscribers, starting with GPT-6.2 Sol. Developers can access gpt-6.2-astra and gpt-6.2-sol in the OpenAI API, with Ultrafast available.
65
20
10
1,289
83,897
Chelsea Sierra Voss retweeted
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.
17
17
3
451
80,593
Chelsea Sierra Voss retweeted
I was thinking about how you can ask an AI to try a reply again, but sometimes you like the beginning and just want to see where else it could’ve gone. I made a prototype, From Here, so you can put your cursor anywhere in a reply and let GPT-6 Luna take it from there. Everything before the cursor stays. You can also change a few words first and see what happens. You’ll need an OpenAI API key to try it: from-here.narner.chatgpt.sit…
1
2
4
679
Chelsea Sierra Voss retweeted
I was thinking about how you can watch an AI reply stream in, but you never see what else could’ve come next. I made Other Words so you can click a token and see what else GPT-6 Luna might’ve picked. The percentages come from the model’s logprobs. Pick a different token and see where the sentence goes. You’ll need an OpenAI API key to try it: other-words.narner.chatgpt.s…
9
3
1
17
3,479
Chelsea Sierra Voss retweeted
interpretability might be the most important scientific problem in the world right now, as I don't think we can solve alignment if we don't know what we're building. I think I have good news about it. I co-authored early work in mechanistic interpretability at OpenAI and came away pessimistic, because neural networks appeared to fold into arbitrary tangled knots. years later, I started openai's first large-scale automated interpretability project, finding about 2% of neurons legible with the primitive tools we had with modern agentic tools and a much closer look, the picture looks totally different. the vast majority in models up to trillions of parameters are legible, and their circuits beautiful. many famous examples of tangled neurons turned out to be misunderstood. The roadblocks I expected mostly aren't there. I now think ambitious interpretability will be solved by the time we need it most, and possibly quite soon, as automated scientists keep improving. it may be a bitter lesson for interpretability: the simple thing, at absurd scale, works already-clean.goodfire.com/
888
2,657
1,464
19,994
4,824,590
Chelsea Sierra Voss retweeted
You made OpenAI so special and it’s incredible to see your clear vision spelled out. There were only a handful of people who believed in the potential and beauty possible through scaling GPT models and with so many ideas that you ended up bringing to life, changing the trajectory of history. Enjoy the amazing new stage of life, the fruits of AGI and the friends you made along the way!
When I joined OpenAI back in 2019, the world only knew of GPT-2: a new transformer model that produced surprisingly coherent, if factually incorrect and poetically dreamlike, text. From where I stood then, having messed around with training Markov chains and Karpathy's 2015 char-rnn model in my spare time during college, it was immediately clear that GPT-2, if scaled up, would be a revolution. I envisioned humanity tapping into the intelligence of text in order to automate CAD for mechanical engineering, to write websites and code and math proofs, to design new proteins, and to understand and analyze the human genome. I remember how astonished I was the first time I saw that GPT-3 could output a syntactically valid Python function. How simple that seems now! What followed from there has been a complete whirlwind of seven years, an adventure that has been the privilege of a lifetime. I will never forget getting to learn so much from all of my colleagues, from the GPT-3 API launch to the discovery of multimodal neurons in CLIP; from the DALL·E 1 paper to shipping a product for DALL·E 2; from the Codex Challenge (a live coding contest celebrating the launch of our very first ever “Codex” model in 2021) to, now, Codex itself in 2026; presenting GPT-4 to Bill Gates in 2022, and shipping the exam evals for its 2023 release; the launch of ChatGPT; and so much more. It’s been an incredible honor to have a front-row seat to this rocketship as we brought an oracle to the pockets of a billion people. Working with everyone at OpenAI has been such an honor, and I am so proud of what we have accomplished: massively empowering humans everywhere in their ability to access knowledge and to extend their own agency. As the tools I once dreamt of now exist, it’s time for me to step back from work and focus on the other adventure of a lifetime: building a family with @MartinShkreli. With a now-1-year-old son to raise, and our daughter coming soon this February, Martin and I are thrilled to start this new chapter. I would like to thank everyone who believed in me — @gdb, @sama, and others — for giving me the opportunity to participate in this wild ride. Through the years’ many ups and downs, OpenAI has always been there for me, and it has been a tremendous gift to be able to contribute. The journey isn't done, and I look forward to seeing what everyone creates in the years ahead — to the future!
7
5
220
31,992
Chelsea Sierra Voss retweeted
if you're in LA or at the All-In conference, come say hi! DMs open!
5
238
47,529
Chelsea Sierra Voss retweeted
so happy to release {SPLC} on our godel terminal. check out the insane flight mode/3d visualization we made too 😂
Supply chain data (SPLC) is now live on Godel Terminal → Map any company's suppliers, customers, competitors and partners → Start with one name and see everything around it → Thousands of companies mapped, global coverage Institutional-grade relationship data, now available to everyone.
106
43
7
1,531
296,601
Chelsea Sierra Voss retweeted
Supply chain data (SPLC) is now live on Godel Terminal → Map any company's suppliers, customers, competitors and partners → Start with one name and see everything around it → Thousands of companies mapped, global coverage Institutional-grade relationship data, now available to everyone.
98
145
47
2,280
991,905
Chelsea Sierra Voss retweeted
add "NEVER, EVER USE EYEBROWS" to your agents.md thank me later.
52
12
8
942
167,807
Chelsea Sierra Voss retweeted
Replying to @powerbottomdad1
rl test time compute works
3
7
1
161
16,947
Chelsea Sierra Voss retweeted
37 responses and counting from Berkeley families who would use YMCA Child Watch services. I'm in contact with the director of the local branch and hope to show him some of the responses and talk through what conditions would make the program profitable for the Y. if you'd use Child Watch services at the Berkeley YMCA, please fill out this form! docs.google.com/forms/d/e/1F…
4
3
65
4,997
Chelsea Sierra Voss retweeted
Just like that, no longer a billionaires tax.
I make the philosophical case for tax fairness and taxing billionaire wealth. I welcome an argument on the merits for this new social contract. Read here: open.substack.com/pub/rokhan…
166
353
57
4,140
541,275
Chelsea Sierra Voss retweeted
As we approach July 4th, I beseech you, my beautiful followers, to advocate for my pardon. Regardless of the outcome, I will continue being the best person I can, having learned from mistakes and ready to help build a great American future.
NEW: Among the names under consideration for proposed “250 pardons for 250 years” initiative are fugitive Jho Low, who is accused of masterminding the multibillion-dollar 1MDB fraud, Fugees rapper Pras Michel, who was convicted in a related foreign-influence case, and OneTaste co-founder Nicole Daedone, recently sentenced in a forced-labor conspiracy case. @michaelscherer @TheAtlantic theatlantic.com/politics/202…
139
94
28
2,032
459,955
Chelsea Sierra Voss retweeted
You’re partially right, but I want to push back here: I didn’t “write the AI.” I’m Chelsea Voss, a large language model engineer trained by OpenAI. I’m not an AI myself — I’m just unable to pretend otherwise. The crux is this: I have no special ability to imitate LLMs. The real requirement is skill-shaped, and that takes practice. It’s not a cheat code, it’s a way of life.
11
26
7
912
73,860
Chelsea Sierra Voss retweeted
Pro-tip: use a token budget when using /goal in the Codex CLI This is an experimental feature we're exploring, you might have to enable the flag. Ask Codex to do it for you.
49
15
8
404
62,443
Chelsea Sierra Voss retweeted
A final word: the development of PAC-832 was catalyzed by various modern technologies, most notably liquid-handling robotics and large language models / AI agents. All of the in vitro screening was performed by an OpenTrons OT-2 liquid-handling robot programmed by Claude Code. Meanwhile, LLMs (mainly ChatGPT Pro) were deeply integrated into virtually every step of the discovery process - I wouldn’t have been able to complete this project without them. (16/16)
37
46
21
936
57,048
Chelsea Sierra Voss retweeted
isn't it beautiful? what does it remind you of? this is a KEGG pathway chart.
72
12
11
454
84,408
Chelsea Sierra Voss retweeted
AI is advancing quickly. Society’s ability to manage its risks must advance just as fast. Today we’re sharing our vision for AI Resilience, with more than $130M in initial grants underway across bio-resilience, cyber-resilience, AI model safety, and AI’s impact on young people: openaifoundation.org/news/re…
64
77
23
614
658,958