Jev: The Model That Killed Chat GPT's Core Idea? RLCD Explained

Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.

Original language: English

Original video: Jev: The Model That Killed Chat GPT’s Core Idea? RLCD Explained · Channel: Fahd Mirza · 05:16 · Watch on YouTube

Topics: AI evaluation and benchmarks, AI models, AI safety

Summary

The video introduces Typesafe’s proposed alternative to RLHF, called RLCD, which its founder argues could better support AI systems that make decisions inside software. RLHF trains models using human rankings, while RLCD is described as training for correct outcomes and calibrated confidence scores. Typesafe says its model, Jeff, can return typed decisions with confidence values, and reports speed, cost, and structured-output benefits. The presenter emphasizes that these claims have not been independently verified and the model is not yet available, leaving open whether calibrated confidence is a better approach than human approval.

Key points

The video frames the debate as a challenge to RLHF, the training method used in some form by many current models.
RLHF uses human rankings to encourage answers that people prefer.
Typesafe founder Dio Almeida argues that RLHF can encourage name-dropping, overconfidence, and unreliability.
Typesafe describes RLCD as reinforcement learning for calibrated decisions, focused on whether outcomes are correct and on confidence estimates.
The company’s model, called Jeff in the transcript, is presented as producing typed decisions with confidence scores rather than conversational text.
The proposed confidence scores could let software act automatically above a threshold or escalate uncertain cases to a human.
Typesafe claims its model is faster, cheaper, and has a lower structured-output error rate than comparable LLMs.
The presenter cautions that these are company-reported claims, the model is not yet available, and the results have not been independently verified.
The video leaves open whether RLCD is a better foundation for autonomous AI, or whether RLHF’s human-oriented behavior is useful when people remain involved.

Chapters

Start	Chapter	Summary
00:01	A challenge to RLHF	The presenter introduces the claim that a contributor to the technique behind ChatGPT now considers its training method a dead end.
01:36	How RLCD differs	The video explains the alleged flaws of optimizing for human approval and contrasts RLHF with Typesafe’s outcome-focused RLCD approach.
03:10	Jeff’s proposed use and performance	Jeff is presented as a model for software decisions, with confidence scores and company-reported speed, cost, and structured-output claims.
04:13	Unverified claims and open questions	The presenter notes that the model and benchmarks have not been independently tested and asks whether calibrated confidence is preferable to human approval.

Quotes

RLHF, reinforcement learning from human feedback. — 01:04

RLCD reinforcement learning for calibrated decisions. — 02:06

It’s a model built to be called by software, not talked to by people. — 03:10

not independently verified model is not out yet. — 04:13

Concepts and entities

RLHF, RLCD, Jev, TypeSafe, ChatGPT, Diego Almeida

Full transcript (original language)

00:01 Jet GPT changed the world forever. It didn’t just launch a product. It reset what an entire industry and honestly what most people think AI is even for. So when someone who helped build the technique behind it, stands up and says, “I think we got it wrong.” That’s worth paying attention to. This video is us trying to figure out is this the next Jet Jeopardy moment or just noise. Let’s find out. This is Fahad Miza and I

00:32 welcome you to the channel. There is a tweet blowing up right now from that exact person, someone who says he helped invent the technique behind Chat GPT. And today he’s telling the world he thinks it was a dead end, not the product, the training method itself. And he’s not just criticizing it. He spent last two years in stealth building the opposite. His name is Dio Almeida and his company is Typesafe. This video is

01:04 not sponsored by this company at all by the way. And to understand what he built, you first need to understand what he is walking away from. RLHF, reinforcement learning from human feedback. That’s the method where a raw language model gets fine-tuned by having humans rank its answers. So the model learns to produce responses people like. This is why chat GPT sounds helpful and agreeable instead of just predicting the next word of the internet. Almost every

01:36 other model you see today runs on this idea some form or another. Diego’s claim is that optimizing a model to please a human judge quietly bakes in specific flaws, he names modropping, overconfidence, and unreliability. Models trained this way, he argues, are built to sign sound right to a person in the loop, not to be right on their own. That’s fine for a chatbot. It’s a real

02:06 problem the moment you want a model making decisions inside software with nobody watching and that is where the real story starts. So instead type safe built what they call RLCD reinforcement learning for calibrated decisions. Look at these side by side and split is obvious. RLHF trains on humans ranking answers chasing approvals and it outputs conversational text with overconfidence baked in as a cost. RLCD trains on

02:40 whether the outcome was actually correct chasing and accurate confidence score and it outputs the typed decisions with a probability attached. Same reinforcement learning backbone underneath both completely different things they are optimizing for. And then the story you know moves forward. The difference you can see shows up directly in this output. And RLHF chat model gives you a sentence take it or leave it. Their model called

03:10 Jeff gives you a typed decision plus a calibrated confidence score. So your code can say act automatically above 90% confidence escalate to a human below that. It’s a model built to be called by software, not talked to by people. And that’s a major difference. And because it’s not generating conversational text token by token, they claim it radically faster and cheaper for those decisions style leaks. Their

03:42 own demo shows roughly a tenth of second and fraction of a cent versus several seconds and over a cent for comparable LLM call. They have shared some of the benchmarks which pushed the arguments further. A claimed 0% structured output error rate versus low single digit for GPT and claude models and an accuracy per dollar chart where Jeff sits well above everyone else cost for similar accuracy.

04:13 All of this worth remembering is type safe’s own reporting not independently verified model is not out yet. There is a very long wait list I guess but nothing has been tested, nothing has been proven. It’s all at the moment words on their blog post, not even independently verified. But anyway, look, the real question is not that is Jeff fast. It’s whether optimizing for calibrated confidence instead of human approval is actually a better foundation

04:45 for AI that acts on its own or whether RLHF’s flaws are a feature, not a bug for anything that still needs a human in the loop. So, let me know your thoughts in the comments as very keen to know them. You guys always add value. So as soon as I get my access to their model, I will be making another video and we are going to really really stretch it to its seams to see how the cookie is crumbling. That’s it. Let me know your thoughts in the comments. Thank you for

05:16 all the