people
the founders, the bench, and the backers — who they are, what they wrote, and what they claim.
the founders
- diogo almeida · ceo · completeskepticcofounder and ceo. google brain researcher 2017–2018, then openai mts january 2020 to january 2024 — co-author on the instructgpt paper (neurips 2022) and the gpt-4 technical report; the company bio says he 'co-invented rlhf and instructgpt.' earlier openai work includes 'a generalizable approach to learning optimizers' with winter, tang, and zaremba. 2018–2020 and education are undocumented in public sources — many same-name false positives. handles: x and hn @completeskeptic, completeskeptic.com. team-page fun fact: top-30 league of legends na, top-2 hearthstone.
- sasha sheng · co-foundercofounder and coo since may 2024. ex-meta/fair research engineer — facebook ios news feed, then news feed backend on content labeling and classification (led the company-wide 'human annotation excellence' conferences), then fair research on vqa and extensive contributions to the open-sourced multimodal framework mmf. first author on neurips 2021 'human-adversarial visual question answering' (with singh, goswami, parikh, kiela, others); co-author on eccv 2022 'mugen.' founded the deeplearners peer-learning community (2017–2019), teaches creative coding at gray area, won seven hackathons in 2023. bs mechanical engineering, shanghai jiao tong; stanford certificate in innovation and entrepreneurship.
- erik (spock) gafni · co-foundercofounder and cto. harvard medical school lab for personalized medicine (senior research associate, ngs methods), then senior bioinformatician at invitae 2013–2017 — main developer and first patent author on its variant-calling system; 'my code helped millions of patients.' senior computational biologist at freenome 2017–2018, then co-founder and principal ml scientist at ravel biotechnologies (oct 2018–mar 2021, $9.5m raised, early-detection liquid-biopsy work, a 2022 scientific reports paper). post-ravel he ran an ml research-engineering consultancy across llms, voice cloning, diffusion, finance, rl, and genetics. two bachelor's degrees (cs + physiology); 44th at us chess nationals in elementary school.
the writing
the corpus is the ceo's — the complete skeptic essays and the three september posts pre-loaded the launch's arguments, and the launch post itself carries the caveats. behind it: the instructgpt paper for diogo, adversarial-vqa research for sasha, and a diagnostics paper for erik.
- 2021-06a generalizable approach to learning optimizers · document, arxiv · diogo almeidathe pre-instructgpt openai paper with clemens winter, jie tang, and wojciech zaremba — learned optimizers trained to generalize across tasks.
- 2022-03the instructgpt paper · document, arxiv / neurips · diogo almeida'training language models to follow instructions with human feedback' — the instructgpt paper (neurips 2022) that grounds the 'co-invented chatgpt and rlhf' credit: a >100x smaller model trained on the right task beat gpt-3.neurips 2022
- 2021human-adversarial visual question answering · document, neurips · sasha shengsheng's first-author neurips 2021 paper — adversarial vqa built to break models, plus eccv 2022 'mugen.' the fair-era research record behind the coo seat.neurips 2021
- 2022-05-19evaluation of cfdna as an early detection assay · document, scientific reports · erik (spock) gafnithe ravel-era paper — 'evaluation of cfdna as an early detection assay for dense tissue breast cancer' (gafni, boley; university of missouri collaboration). the cto's publication record is in diagnostics, not llms.
- 2026-06-19ai: too good to be true, too bad to be useful · post, typesafe blog · diogo almeidathe first typesafe blog post — a video-format page whose tl;dr card argues rlhf-trained language models 'please humans and assist rather than make reliable autonomous decisions.' the thesis in embryo, three months before launch.
- 2026-07-04scaling laws, honestly · essay, complete skeptic · diogo almeidathe complete skeptic opens — 'sane takes in an insane world' — with a corrective history of the scaling-laws literature.
the original scaling laws were wrong due to a bug
- 2026-07-24is it even possible for the chinese labs to distill us models? · essay, complete skeptic · diogo almeidathe second complete skeptic essay — whether chinese labs can distill us models — lab-competition analysis ahead of the launch.
- 2026-09-03the wrong path post · post, linkedin · diogo almeidathe linkedin post that opened launch fortnight: 'in 2020, nobody at openai dreaming of agi imagined we'd stop at chatbots.'
at openai, i co-invented chatgpt and rlhf. unfortunately, i think they've taken ai down the wrong path
when models are trained to serve humans, they would rather hallucinate real-looking answers than disappoint you
- 2026-09-09(kv) cache rules everything around me · essay, complete skeptic / typesafe blog · diogo almeida'(kv) cache rules everything around me' — the economics-of-inference essay, first of the three launch-week pieces.
the biggest cost to devs of an agent is re-reading its own context, which costs the provider close to nothing
- 2026-09-10the bitterest lesson · essay, typesafe blog / complete skeptic · diogo almeidasutton's bitter lesson pushed one level up — the instructgpt anecdote as evidence: 'gpt-3 was an incredible model trained to predict the next token… but people wanted something that followed instructions more than they wanted a super-powered autocomplete.'
doing the right task > data > compute > algorithms
you get what you optimize for and the bitterest lesson in ml is that the most important part of it isn't ml at all
- 2026-09-11lies, damned lies, and benchmarks · essay, typesafe blog / complete skeptic · diogo almeida'lies, damned lies, and benchmarks' — 'many of the spikes in jagged intelligence are the benchmarks'; 'no standard benchmark table in our model releases.' the essay that made the workflow-evals choice legible — and the one critics turned back on the launch.
i'm not arguing against benchmarks… i'm against benchmaxxing
benchmarks are a shortcut to get around the hard problem of trust… a loan
- 2026-09-15introducing system one models & jev · essay, typesafe blog · diogo almeida'introducing system one models & jev' — the launch post: 'think of jev as a frontier-intelligence function call'; choice, score, and noul primitives; rlcd; the jevons naming faq; and an unusually caveat-heavy claim table.
models have been superhuman at chat for years, so where is all the automation? this has been my driving question for the last four years
we started typesafe because we believe that ai needs an interface software could depend on
hn #1 — 1,790 points, 473 comments - 2026-09composable ai: build prod, not god · document, typesafe.ai · diogo almeida'composable ai: build prod, not god' — the unsigned company manifesto carrying the ceo thesis: 'the shortest path to an ai-based economic revolution,' operationalized in a footnote as global tfp growth reaching 3% within five years and holding for ten.
we're building prod, not god
intelligence today is like databases before sql
- 2026-09-15the jev launch thread · thread, x · diogo almeidathe x launch thread with video — '20-200x faster • 40-400x cheaper… afict the shortest path to ai-based economic revolution.'
after co-inventing chatgpt, i kept asking myself: why have superhuman chat models not led to agi?
~4.21m views, 19.9k likes per latent space's capture - 2026-09-15the ceo in the comments · thread, hacker news · diogo almeidadozens of substantive ceo replies in the launch thread — accepted the 'zero-shot classifier' reframing ('exactly right!'), conceded the confident-error case, held the architecture back. 'i do think that subconscious is not only much smarter than we give it credit for, but also much more robust than the jagged frontier of current llms.'
that's right, but because these models are probabilistic, it's also possible to be confidently wrong
architecture is close to the chest for now, but we have talked about writing a paper
the team
- the engineering benchlinkedin lists 23 employees (+475% yoy), mostly us plus serbia and romania. named members: elise liu (head of product / mts, 'structured large programmable models'), eunji jinny seo (founding product designer), allie laabs (founding devrel; ex-2k games, trion, solarcity, autodesk), clayton ward (founding engineer; autodesk, amazon, sig), noam samuel (backend/api; stripe, square), carl you (full-stack ai/ml platform; airbnb, plaid), nathan leclaire (mts feb 2026–; resolve ai, bauplan, honeycomb, docker; hn's 'zenlikethat'), eugene shvarts (mts apr 2025–; anthropic product researcher, teleport head of data science, uc davis applied math phd 2023), ted kalaw (mts mar 2025–; meta 2013–2016, ripple), dmitry norkin (mts mar 2026–; wayvee analytics), daniel gafni (mts apr 2026–; dagster labs pipes, anam.ai; shares the cto's surname — relationship unconfirmed), jeffrey lin and evin sellin (mts), kevin zhang (hiring posts).
- the backersthe $40m seed was led by dcvc — gp james hardiman is quoted in the press release — with forbes reporting a ~$200m post-money via a single 'person familiar.' the aws generative ai accelerator (2024) is the only other named backer, a credits program rather than equity. the site says only 'backed by top-tier investors'; no other investors or board members are disclosed.
interviews and talks
- 2026-07-01what's next after rlhf? · ai engineer world's fairthe first public articulation of the anti-rlhf thesis — 'more accurately, i think this should be called what's next after the chatgpt era… and my hint for you guys is it is not the claude code era.' video published july 31.
- 2026-09-15this $200 million startup wants to fix ai's overconfidence problem · forbesrashi shrivastava's launch-day interview for forbes' the prompt — 'you can actually automate something instead of automating nothing because the llms are too unreliable as an aggregate'; broke the ~$200m valuation.
the philosophy
themes tracked across the writing: automation over assistance, the rlhf critique, calibration and epistemic honesty, benchmark skepticism, composable intelligence, the economics of cheap intelligence, prod, not god.
- 2026the co-inventor creditthe 'co-inventor' credit — the company bio and linkedin post say diogo 'co-invented chatgpt and rlhf.' the documented record is co-authorship on the instructgpt paper (neurips 2022), the gpt-4 technical report, and acknowledged contribution to chatgpt; rlhf as a method predates instructgpt (christiano et al., 2017). coverage varies between 'co-author' and 'co-inventor.'treat 'co-inventor of rlhf' as the company's own phrasing — the instructgpt authorship is solid; the stronger claim is marketing grammar, not a paper trail.
- 2018–2020the missing yearsthe thin public record — diogo's education and his 2018–2020 employment are not in public sources; linkedin shows google brain 2017–2018 then openai 2020–2024. same-name false positives complicated the search. daniel gafni shares the cto's surname — any relationship is unconfirmed and should not be asserted.recorded as a gap rather than filled — nothing in the corpus contradicts the linkedin spine, but two years are simply unaccounted.
- 2026-09the real wardrobethe launch-video suspicion — hn commenters guessed the launch video was ai-generated from outfit changes. diogo denied it in-thread: 'definitely not ai-generated - this is my real wardrobe… real voice actor but slightly sped up.'resolved-by-denial — a minor texture item, kept because it documents how closely the launch was scrutinized.
AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.