Jev introduces a new shape of LLM

Join the discussion

• Hacker News
• Reinforcement Learning for Calibrated Decisions (RLCD)

  What is a Calibrated Decision? Under my definition, calibration is the search for errors in measurement. And a separate decision rule is applied to that measurement.

  I see little rigor and precision in the thinking and description around what this model does, much of what I see looks like slop these days...

  by acyou
• I actually find the name System-1 as a nod to Daniel Kahneman’s Thinking fast and slow book kinda nice. It’s an interesting analogy
  by syntaxing
• I thought this wasn't really an LLM sure text as input but the folks at Typesafe have been very adamant this is a new class of model not just another LLM.

  Although to be fair we don't know enough about the architecture.

  by allanmacgregor
• Any idea how they get the probabilities? The probabilities themselves are estimates so knowing the tightness of their distributions will be helpful
  by malshe
• I think "Black boxes are back in fashion" is missing the point. I think LLMs are still largely black boxes, and I don't think chain of thought is representative of any degree of inner machination. Asking it questions to justify itself is at best a facsimile, and for the most part it's useful, but it's fundamentally a facsimile.

  Where I understand Jev to be a significant jump is that afaik the confidence scoring is actually derived from the normalised probabilities, and not a continuation in a chain of prediction masquerading as "confidence."

  by hresvelgr
• About jev being a black box and the potential for bias, I think it boils down to what questions you are asking the model.

  Broad questions like Is this resume good / score this city will ofcourse be biased but I think jev encourages more granular focused questions like Score this candidates Python experience / Rate this city for its food which then allows you to introduce your own biases in which questions you ask and how you combine their answers.

  In this way I think jev like models can be easier to reason about for critical decisions.

  by aszen
• I’m not sure I understand the hype around this model. Isn’t this just an llm with a chat template, with the options prefix cached?

    <option>option A</option> <option>option B</option><endofoptions>userprompt<eos>
  

  Then the llm is constrained to a few special tokens indicating the possibilities? e.g. <option1> <option2>
  by tipsytoad
• If you're looking for "Hot Dog/Not A Hotdog" type (i.e., heavily bounded) answers from LLMs, make certain to at least consider (or better still benchmark) traditional AI and machine learning approaches against LLMs. The traditional approach is almost certainly less computationally expensive and might work better for your problem than an LLM. I recall hearing stories from people on a popular podcast, that they benchmarked and went with the traditional approach because it was more accurate, deterministic, and cheaper. Many had always wanted to try traditional AI/ML, but only got the go ahead recently to add AI to their project from management because of the AI/LLM hype. They were surprised to find that when they benchmarked, the traditional approaches outperformed LLMs both in terms of accuracy and cost. Don't get fixated on LLMs to the point of ignoring older tools that might do a better job.
  by infamia