This new AI model isn't an LLM. It's 20x cheaper and plays Doom in real time (Jev explained)

節目時間軸

1. 0:00

   TypeSafe Labs launches JEV, a new class of System 1 model, after two years in stealth.

   • TypeSafe Labs, co-founded by Diogo Almeida (a co-creator of RLHF and ChatGPT), emerged from stealth after two years to launch JEV, a new class of model they call a System 1 model.
   • The name System 1 comes from Daniel Kahneman's Thinking Fast and Slow, referring to fast intuition rather than the slow deliberative reasoning of LLMs.
   • JEV is priced at roughly $42 per billion input tokens, with output tokens considered 'too cheap to meter,' a pricing model that is orders of magnitude cheaper than typical LLM pricing.
2. 3:02

   Alex demos a real-time JEV-powered tweet classifier running on his X timeline.

   • Alex built a browser extension that scores every tweet on his timeline with JEV in real time, processing about 12 tweets per second at essentially zero cost.
   • JEV is 4-5 times faster than a fast model like Qwen running on Cerebras LPU hardware and about 20x cheaper, enabling a 'real-time cognitive firewall' for social feeds.
   • Because JEV returns probabilities rather than generated text, categories appear before an LLM could even begin responding, enabling use cases like semantic ad blocking.
3. 6:22

   Ali explains the core innovation: machine-to-machine AI without a human interface translation layer.

   • Ali argues that using LLMs to call tools is wasteful because it translates machine code into human interface language and back again, like using a bipedal robot instead of a dishwasher.
   • JEV exposes latent AI intelligence as machine-readable, type-safe probabilities and typed objects, enabling machine-to-machine use without ever returning to user space.
   • The training approach is called reinforcement learning for calibrated decisions (RLCD), doing parallel processing of all questions at once rather than generating text character by character.

關鍵概念

• System 1 model— TypeSafe's new class of model inspired by Kahneman's fast-thinking concept, designed for real-time decision making rather than text generation.
• JEV— The first public System 1 model from TypeSafe Labs that outputs calibrated probabilities instead of generating text.
• TypeSafe Labs— The stealth startup behind JEV, co-founded by Diogo Almeida, a co-creator of RLHF and ChatGPT.

精選金句

• I think that Jev can categorize all of my life's work in terms of the text that I outputted as a human with less than 40 dollars.

  🤯

  — A lifetime of human writing can be processed by JEV for under $40, revealing an almost absurd gap between human effort and machine cost.

  跳至 0:05

• It's so weird to me that we are using these, like, big old language models to create the text or the instructions to interact with an interface that was already on the other side of the interface is machine code.

  💡

  — Overturns the assumption that LLMs are the natural way to automate tasks, exposing the wasteful translation layer between human language and machine code.

  跳至 7:48

可執行的洞察