DisputedFundraising

TypeSafe debuts with a $40 million seed round as Diogo Almeida doubts his own Jev model

Protos reports TypeSafe launched with a $40 million seed round and the Jev AI model, while founder Diogo Almeida says the RLHF work behind it 'is not useful'.

Why it matters

A $40 million seed round is backing a startup whose founder markets himself on a ChatGPT credit that the 2022 InstructGPT paper does not support, and whose model he describes as ineffective. Because only Protos has reported all of this, the round, the model and the authorship dispute rest on a single outlet.

Where the sources disagree

Almeida's role in building ChatGPT

TypeSafe pitch: solved the AI hallucination problem

Protos

Almeida: RLHF 'is not useful' and the invention does not work well

Protos

Almeida presents himself as a ChatGPT 'co-creator'; Protos says the 2022 InstructGPT paper lists him fourth of 20 authors and its footnote marks him as one of nine 'primary' authors rather than a team lead.

Key facts

TypeSafe emerged with a $40 million seed round.

40,000,000 USDReported by Protos

TypeSafe's pitch claimed the startup had solved the AI hallucination problem, a problem Almeida says was created during his time working on ChatGPT.

Disputed figureReported by Protos

TypeSafe is marketing a new AI model called Jev built on Reinforcement Learning from Human Feedback.

Reported by Protos

Almeida markets himself as a ChatGPT 'co-creator', and Protos reports the 2022 InstructGPT paper that applied RLHF to language models lists him fourth of 20 authors.

Disputed figureReported by Protos

A footnote in the InstructGPT paper marks Almeida as one of nine 'primary' authors rather than a team lead.

Reported by Protos

Protos states that RLHF predates the 2022 InstructGPT paper by nearly five years and that its foundational method came from six other researchers.

Reported by Protos

Almeida says RLHF 'is not useful' and calls ChatGPT 'unreliable without human supervision'.

Disputed figureReported by Protos

Almeida blames RLHF for 'mode dropping, overconfidence, and an overall lack of', a list that the published text leaves unfinished.

Reported by Protos

What happened

Protos reports that Diogo Almeida launched his startup TypeSafe this month with a $40 million seed round and a new AI model called Jev, which the startup pitched as an answer to the hallucinations problem Almeida says arose during his time working on ChatGPT. The campaign calls Almeida a ChatGPT 'co-creator', and Protos disputes that: the 2022 InstructGPT paper that applied RLHF to language models lists him fourth of 20 authors, and a footnote marks him as one of nine 'primary' authors rather than a team lead. Protos adds that RLHF predates that paper by nearly five years and that its foundational method came from six other researchers. In the same campaign, Almeida says RLHF 'is not useful', calls ChatGPT 'unreliable without human supervision' and blames RLHF for 'mode dropping, overconfidence, and an overall lack of', a list the published text cuts off mid-sentence. TypeSafe's pitch and Almeida's own account therefore do not agree on whether the model works, and no outlet other than Protos has reported the round, the model or the authorship dispute.

How the story developed

  1. First report

    First report by Protos

    Meet Jev, the AI that claims it can’t hallucinate

    Protos

  2. Update

    TypeSafe emerges with a $40 million seed round and the Jev model

    Protos reports the startup and its AI model appeared this month; no round date is given beyond that.

    Protos

  3. Official statement

    Almeida calls RLHF 'not useful' and ChatGPT 'unreliable without human supervision'

    The founder's own statements undercut the pitch his startup is running.

    Protos

  4. Update

    Protos publishes its account of Jev and TypeSafe

    The report challenges the 'co-created ChatGPT' claim and cites the InstructGPT paper's author list and footnote.

    Protos

Affected entities

Original sources

PublisherReportRolePublished
Protos
Crypto media
Meet Jev, the AI that claims it can’t hallucinateOriginal