history
the company, sourced and dated — the founding, two years of stealth, the argument in public, and the september launch.
the founding2024
an openai researcher leaves, registers typesafe.ai, and quietly starts a lab on a name the scala company vacated — this is not the lightbend/akka typesafe.
- 2016-02-23not the old typesafedisambiguation first: the original typesafe inc. — the scala/akka company founded in 2011 by martin odersky, jonas bonér, and paul phillips — renamed itself lightbend on this date and now does business as akka. the new typesafe ai is unrelated: different domain, field, founders, and era. its name refers to models emitting type-safe structured values.
- 2024-01almeida leaves openaidiogo almeida leaves openai — mts january 2020 to january 2024, co-author on the instructgpt paper and the gpt-4 technical report — and lists the typesafe ceo role starting january 2024. self-listed on linkedin.
- 2024-05-07typesafe.ai registeredthe domain typesafe.ai is registered through cloudflare — the earliest dated artifact of the new company.
- 2024-05the co-founderssasha sheng lists the typesafe co-founder role starting may 2024 — ex-meta/fair research engineer, first author on neurips 2021 human-adversarial vqa. erik gafni is the third co-founder and cto.
- 2024-05-28github org createdthe github org typesafe-ai is created — dormant until the 2025 forks and the august 2026 repo wave.
- 2024-09-13the aws acceleratortypesafe ai is named to the 2024 aws generative ai accelerator — 80 startups picked from 4,700+ applications, up to $1m in aws credits each. aws's blurb: 'building a new type of general frontier model for reliable ai systems to automate real work and decision making.' cb insights logs the round 2024-09-17.
the stealth years2024–2026
two years of near-silence: a homepage titled 'intelligence beyond chat,' a slow team build, soc 2, and github forks that hint at non-autoregressive research.
- 2025-03the first hiresthe earliest known non-founder hires: ted kalaw joins as mts in march 2025 (ex-meta, ripple) and eugene shvarts in april 2025 (ex-anthropic product researcher, uc davis applied math phd). the stealth homepage 'intelligence beyond chat' carries a growing team roster.
- 2025-05the telling forksthe github org forks vllm in may 2025 and llada — a diffusion-lm implementation — in july 2025. consistent with non-autoregressive model research; the company has never confirmed the connection.
- 2026-01@typesafeaithe @typesafeai x account is created — the company's first public surface outside its own site.
- 2026-02the 2026 buildoutthe 2026 buildout: nathan leclaire joins as mts in february (ex-docker, bauplan, resolve ai), dmitry norkin in march, daniel gafni in april (dagster labs; shares a surname with the cto — relationship unconfirmed), and eunji jinny seo as founding product designer in july. linkedin's count: 23 employees, +475% yoy.
- 2026-07soc 2typesafe becomes soc 2 compliant, per nathan leclaire's linkedin — enterprise-readiness groundwork laid while still in stealth.
the argument in public2026
the argument goes public before the product does — teasers, the 'what's next after rlhf?' talk, the complete skeptic essays, and three september essays that pre-load the launch.
- 2026-04-03the first teaserthe first public teaser at a 'founders you should know' event — per allie laabs's repost of diogo's linkedin: 'don't tell big ai, but we finally gave a teaser to what we're doing.'
- 2026-04where is all the automationdiogo gives a talk at an ai council conference: 'llm history, mystery, and optimization to figure out where the f**k is all the automation' — the driving question in public, five months early.
- 2026-06-19too good to be true, too bad to be usefulthe first typesafe blog post: 'ai: too good to be true, too bad to be useful' — a video-format page whose tl;dr card argues rlhf-trained models 'please humans and assist rather than make reliable autonomous decisions.' the earliest dated artifact on the site.
- 2026-07-01what's next after rlhf?diogo speaks at ai engineer world's fair, track 9: 'what's next after rlhf?' — the first public articulation of the anti-rlhf thesis. 'the team i was part of basically invented post-training as a concept'; 'i'm one of the few people at openai who actually hates on chatgpt.' video published july 31; latent space later disclosed typesafe previewed at an aie event ~august.
- 2026-07-04complete skepticdiogo launches the complete skeptic — 'sane takes in an insane world.' early posts: 'scaling laws, honestly' ('the original scaling laws were wrong due to a bug') and 'is it even possible for the chinese labs to distill us models?'
- 2026-08-08the repos surfacethe public repos begin appearing: system-one-adapter-python (aug 8), skills (aug 24), overwatch (aug 31), and the python + js sdks (sep 4) — the launch infrastructure landing in the open weeks early.
- 2026-09-03the wrong path postdiogo on linkedin: 'at openai, i co-invented chatgpt and rlhf. unfortunately, i think they've taken ai down the wrong path' — models 'trained to serve humans… would rather hallucinate real-looking answers than disappoint you.'
- 2026-09-09the september essaysthree essays in three days load the launch: '(kv) cache rules everything around me' (sep 9), 'the bitterest lesson' (sep 10 — 'doing the right task > data > compute > algorithms'), and 'lies, damned lies, and benchmarks' (sep 11 — benchmarks as 'a loan' of trust; 'no standard benchmark table in our model releases').
the launch2026-09
out of stealth with jev, a $40m dcvc-led seed, and the top of hacker news — then 48 hours of grilling, independent evals, and an open-source reimplementation.
- 2026-09-15out of stealth, $40m seedout of stealth. businesswire, 3:00pm edt: $40m seed led by dcvc (gp james hardiman quoted). diogo's launch post introduces 'system one models' and jev — unstructured state in, typed probabilistic decisions out; choice, score, and noul primitives; rlcd training; claimed 70–500ms latency and $42 per billion input tokens with output unmetered (all self-reported). forbes reports a ~$200m valuation citing a person familiar — single-source.
- 2026-09-15the hn megathreadthe launch thread goes to #1 on hacker news — 1,790 points, 473 comments, submitted by 'albelfio' (affiliation unknown); the original title 'jev: new frontier model 40-400x cheaper and 20-200x faster' is changed within the hour after misleading-title complaints. diogo answers dozens of comments as completeskeptic. his x thread reaches ~4.2m views per latent space's capture. a discord town-hall ama runs the same evening.
- 2026-09-15the same-day evalsthe first independent evals land same-day from early-access users: every's mike taylor runs 777 judgments across 37 docs in under 0.7s for about a quarter of a cent; the every ceo's planted-defect test has jev catch 6 of 7 vs fable 5.1's 7 of 7, ~25x faster at ~1/580th the cost. good start labs logs 91.5% jev↔fable agreement on 6,003 rubric checks at $160 vs ~$33,000 per million graded answers.
- 2026-09-16the press wavethe press wave: the register ('jev is hallucination-free… really isn't a fair comparison'), the decoder (eval-methodology limits), siliconangle, tech startups, ai news, gadgetsnow — plus two techmeme clusters and latent space's title story. jev lands on the vercel ai gateway.
- 2026-09-16the skeptic ledgerthe skeptical ledger forms: orcarouter separates vendor from independent numbers — jev 'wins the cost and latency columns… and loses the accuracy column' (67.8% vs 74.1% on the vendor dashboard); daily texas news documents three inconsistent speed claims, the non-empirical 0% hallucination footnote, and the reasoning-disabled wikiracing comparison. near here's moderation test is the most favorable: 96% vs mistral's 84%.
- 2026-09-16jevlikethe replication starts inside 48 hours: vinnylarouge/jevlike, an mit-licensed option-attention reimplementation with the same i/o shape, posts doom and chess demos — 69 points on hn alongside a satirical two-hour 'qwen-2.5-1b-rlcd' clone video.
AI-drafted at Ben Guo's direct request and credited to Hraness; every claim links to its cataloged source.