Talks & appearances

What’s next after RLHF?

Diogo Almeida, CEO · AI Engineer

Almeida’s argument: RLHF made models extraordinary at pleasing a human in the loop. Dependable automation needs a different objective — correct, calibrated decisions software can use without continual supervision. Assistance and automation pull in different directions in optimization space. TypeSafe’s third path is RLCD / calibrated decisions for software, not another assistant UI.

What to take from it

• RLHF is great at pleasing a human in the loop — that is a different job from running software.
• Assistance and automation pull in different directions in optimization space.
• TypeSafe’s stated third path is RLCD: calibrated decisions for code, not another chat UI.

Official talk page with transcript. Almeida places himself between making ChatGPT and criticizing the objective behind it.

Original page ↗

AI: too good to be true, too bad to be useful

Diogo Almeida, CEO · AI Council

Why some AI apps over-promise and under-deliver. Almeida looks at LLM history and optimization and asks where the automation actually went. TypeSafe’s matching blog URL (Jun 19, 2026) is a YouTube embed of this talk, not a prose essay.

What to take from it

• The talk is the source — TypeSafe’s blog URL for this title is a YouTube embed, not an essay.
• The question is where the automation went after models got superhuman at chat.
• Watch it if you want Almeida’s own telling; the 101 stays on the written path.

Watch the talk ↗ · Original page ↗

Founders You Should Know

Diogo Almeida, CEO · Founders You Should Know · San Francisco

In-person SF showcase. TypeSafe’s LinkedIn announced the talk: if the last generation learned to talk to humans, what does the next generation need to look like to run real systems? Recording titled “Diogo Almeida - TypeSafe AI.”

What to take from it

• This predates the September Jev launch — useful for the thesis, not the product FAQ.
• The framing: last generation talked to humans; next generation has to run systems.

Announcement: linkedin.com/company/typesafe-ai.

Original page ↗

Jev explained live — Choice, Score, Noul

Allie Laabs, founding DevRel · ThursdAI with Alex Volkov

Two days after launch, Laabs walks through System One primitives. Her line: tool-calling LLMs translate machine work into a human-shaped interface, then back again — “an extremely expensive translation layer.” Show notes also cover a live tweet-classifier demo and TypeSafe’s published latency/price claims (company numbers).

What to take from it

• Choice, Score, and Noul are the live primitives — the same ones this 101 starts with.
• Laabs’s line: tool-calling LLMs are an expensive translation layer between humans and machines.
• Latency and price mentioned on the show are TypeSafe claims, not this guide’s measurements.

Secondary show notes.

Original page ↗

Discord town hall / AMA

Diogo Almeida · Allie Laabs · TypeSafe Discord

Official X announced a launch-day AMA (Sep 15) and a follow-up town hall (Sep 16). Allie posted a Doom-demo walkthrough for Sep 18. Discord is the live venue; this 101 links the public announcements, not a private archive.

What to take from it

• Launch week had a Sep 15 AMA and a Sep 16 town hall — announced on official X.
• This 101 does not republish Discord chat. The public announcements are the receipts.

Original page ↗

A new kind of AI model from a ChatGPT inventor

Diogo Almeida interviewed by Tim Fernholz · TechCrunch

Secondary press. Almeida: “We have lightning in a bottle, and yet it is not useful… we are optimizing for human language… computers speak a different language.” Also quotes Armin Ronacher on 50% vs 95% probability, and developer anecdotes — label those as press, not TypeSafe metrics.

What to take from it

• Almeida’s quote: models are optimized for human language; computers speak a different language.
• Developer anecdotes and “thrilling” framing are press — not TypeSafe metrics.

Original page ↗