ChatGPT Co-inventor Launches Decision Model Jev, Attracting Attention
Diogo Almeida, co-inventor of the ChatGPT training mechanism, announced on social platform X that his new company TypeSafe AI has officially released an advanced AI model named Jev after two years of dedicated research and development. Jev employs a novel training method called RLCD (Reinforcement Learning for Calibrated Decisions), focusing on "composable intelligence specifically for decision optimization," rather than traditional text or image generation capabilities.
Almeida stated in his post that Jev is 20 to 200 times faster than leading large language models, with costs reduced by 40 to 400 times (with output tokens even being free), claiming it is the "shortest path to an AI-based economic revolution." According to official test data, Jev's response time for the same task is 0.114 seconds, while OpenAI's GPT-5.6 Terra takes 8.566 seconds for the same task. In terms of pricing, Jev charges only $0.042 per million input tokens with no output fees, compared to GPT-5.6 Terra's pricing of $2 and $12 per million tokens for input and output, respectively. When compared to another flagship model, Fable 5.1, Jev's cost advantage can reach 238 times.
Unlike traditional large language models that predict the next token sequentially, Jev uses a parallel processing architecture to return probability-calibrated structured answers simultaneously through three types of "problem primitives": Choice, Score, and Noul, rather than generating natural language text. Official demonstrations show that Jev can directly read structured state data from the game Doom, outputting probabilistic operational decisions in real-time and completing the entire game with a decision latency of less than 100 milliseconds. The company positions Jev as a "software primitive" that can be repeatedly called and embedded within applications by developers, rather than a consumer-facing chatbot product.
TypeSafe AI was co-founded by Almeida, Erik Gafni, and Sasha Sheng, and has secured $40 million in seed funding led by DCVC. Almeida was previously a core member of OpenAI's training post-processing team and a co-author of the InstructGPT paper, recognized as one of the key inventors of the reinforcement learning from human feedback (RLHF) optimization technique now adopted by nearly all commercialized language models.
Almeida referred to the conversational AI wave initiated by ChatGPT as a "strange detour" from the industry's original automation goals, pointing out that "today's AI is designed to cater to human preferences and provide assistance, fundamentally hindering true automation." He also stated, "No matter how absurdly wrong the model's judgments are, they appear to be correct," and that "over-commitment is a design feature (of the current industry), not a flaw." Currently, Jev has opened an early access application channel (typesafe.ai) and can be directly called through the Vercel platform.
From a funding and industry logic perspective, TypeSafe AI chooses to completely bypass the current giants' concentrated investments in general dialogue and content generation tracks, instead betting on decision automation in this niche scenario. Essentially, it seeks a differentiated capital efficiency path in the face of direct price and computational power competition with leading labs like OpenAI and Anthropic—entering enterprise-level scenarios such as customer service routing, business process automation, large-scale data classification, and AI agent tool invocation, which are highly sensitive to latency and cost, with reasoning costs far below flagship models. The logic behind early investors like DCVC is that once Jev is widely integrated into the "decision layer" of existing AI agent workflows, it will create strong stickiness for the developer ecosystem; for existing applications that rely on high-latency, high-cost flagship models to complete simple judgment tasks, Jev's emergence signifies potential replacement pressure and cost restructuring space.
Source: Public Information
ABAB AI Insight
During his time at OpenAI, Almeida was not an ordinary employee but a co-author of the InstructGPT paper, directly involved in the invention of the reinforcement learning from human feedback (RLHF) training paradigm—this method was later integrated into the GPT series and nearly all mainstream commercial large language models, forming one of the core technological foundations that allowed ChatGPT to be "compliant" and commercially scaled. He subsequently chose to spend two years in stealth development rather than continuing to iterate on the existing conversational AI path, a path that mirrors the behavior of some core members of OpenAI who left to establish their own ventures and bet on different training philosophies after directly participating in the invention and then questioning its limitations.
In terms of funding mobilization, the scale of financing chosen by TypeSafe AI ($40 million seed round) is extremely restrained compared to the tens of billions raised by OpenAI and Anthropic, reflecting its differentiated approach: not competing directly with giants on computational power and parameter scale of general large models, but concentrating funds on a narrower yet higher-frequency scenario—structured decision reasoning. DCVC, as an early institutional investor in deep tech, is betting on this direction, essentially wagering that "the next layer of profit in AI infrastructure" lies not in stronger conversational abilities but in low-cost decision components that can be repeatedly called by massive enterprise applications. This bears structural similarities to the early capital path of cloud computing, which gradually segmented specialized inference chips and databases from general computing resources.
This entrepreneurial logic of "jumping out of the main track to bet on narrower but more essential scenarios" is similar to the paths some companies attempted to find niche cuts in the general large model landscape dominated by giants, but TypeSafe AI's chosen cut is more extreme—directly eliminating text and image generation capabilities while retaining only the single capability dimension of decision-making. In terms of the industry's current stage, the route represented by Jev as a "specialized small high-frequency reasoning model" is at a transitional period where the arms race of general large models has reached a bottleneck, and the market is beginning to seek differentiated infrastructure layer opportunities, rather than a mature and stable competitive landscape.
This phenomenon essentially belongs to "technological substitution": its mechanism occurs because current mainstream large language models bundle "generating natural language" and "completing decision-making" capabilities within the same architecture and pricing system, forcing many high-frequency, low-latency scenarios that only require a binary judgment or scoring result to pay redundant computational and time costs for unused language generation capabilities. TypeSafe AI, through the RLCD training method and parallel structured output architecture, separates and optimizes the "decision" capability from "generation," essentially decoupling the capability bundling of large models—this "decoupling efficiency revolution" could, once validated for commercial viability, push a batch of existing products that also rely on high-latency, high-cost flagship models to complete simple judgment tasks to turn towards procuring more specialized and cheaper decision layer components, thereby reshaping the layered structure of AI reasoning infrastructure.
ABAB News · Cognitive Laws
Speaking AI is expensive, judgment AI is fast.
The more capabilities are bundled, the higher the redundant costs; the more capabilities are decoupled, the more extreme the efficiency.
Giants compete on parameter scale, disruptors compete on single-point efficiency.