Jev decides 20–200x faster than LLMs — AI Digest

Jev decides 20–200x faster than LLMs — AI Digest

TypeSafe released Jev, a model that decides instead of writing text: 20–200x faster with output tokens free. Periodic Labs lifted an internal materials benchmark from 2.7% to 55.3%, and Gemini 3.8 Live took the top spot in speech-to-speech at $0.84 per audio hour.

Today's main stories

Jev: TypeSafe's decision model claims 20–200x speed and free output tokens

TypeSafe released Jev, a model that does not write text at all — it decides: classifying, routing, scoring and picking from a predefined set of options. Launch author Diogo Almeida claims 20–200x faster inference and 40–400x lower cost than small frontier models, with output tokens not billed at all (TypeSafe announcement, The Rundown breakdown). A decision model returns a choice from a fixed set — a class label, a route, a score — instead of free-form text. Jev is trained with RLCD and is non-autoregressive by design, which is where the speed gap comes from: there is no token-by-token generation, just one pass to the answer. Practitioners immediately named the target slot: replace LLM calls wherever the system only ever wanted a structured choice — judges inside pipelines, request routers, scoring (@omarsar0, @chaseleantj).

Periodic Labs' Neon lifts an internal materials benchmark from 2.7% to 55.3%

Periodic Labs unveiled Neon, a model trained in a closed loop with physical laboratories: experiments run continuously and their results feed straight back into training (Liam Fedus announcement, Periodic Labs). The company says it used 1,300 H200 accelerators, months of proprietary experimental data, mid-training followed by RL — and an open base model rather than one built from scratch. A community summary adds the numbers: it starts from Kimi K2.6, lifts success on the internal FrontierXRD eval from 2.7% to 55.3%, and beats GPT-6 Astra and Claude Fable 5.1 at lower inference cost (@brianzhan1). The genuinely unusual part, engineers noted, is RL on real experimental data from physical labs rather than simulation (@vwxyzjn). This is a working template for "AI for science": a proprietary data stream on top of open weights.

Gemini 3.8 Live tops the speech-to-speech ranking at $0.84 per audio hour

Google shipped Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, conversational models that talk, reason and run tasks in the background without breaking the flow of a call (Google DeepMind). What matters to developers: 97 languages, asynchronous tool calls while the model is still speaking, access through the Gemini API and AI Studio, and ready integrations with LiveKit, Pipecat, LangChain and Vercel (@_philschmid breakdown). An independent measurement from Artificial Analysis puts Gemini 3.8 Live Extended Thinking (High) first on its speech-to-speech index at 82.6 against 81.5 for GPT-Live-1 Astra, and first on Tau Voice at 68.6% (Artificial Analysis). Pricing from the same source: $0.84 per hour of input audio for the standard model, $3.50 for Extended Thinking High.

Perplexity built a DynamoDB replacement with two engineers and hundreds of agents

Perplexity says it built and shipped CobbleDB, an in-house DynamoDB replacement for search serving, in two months with two engineers and hundreds of continuously running AI agents (Aravind Srinivas). The company reports median batch-read latency dropping from 31.4 ms to 5.60 ms, p99 from 123 ms to 24.2 ms, and at least 20% savings against DynamoDB (Perplexity). The "hundreds of agents" framing deserves caution, but the case matters for a different reason: the agents were not doing one-shot code generation, they were doing sustained systems work — migration, testing, rollout support. That is a different class of task from editor autocomplete.

Numbers and facts

Different views: will a decision model replace LLM calls in production?

The argument is not about Jev's numbers but about where it applies: a model that physically cannot produce free-form text covers part of the workload completely and the rest not at all. Opinions over 15–16 September 2026 settled into three positions.

For. Practitioners point at an obvious niche: production systems are full of places where only a structured choice is ever wanted, yet full generation is what gets paid for — judges in pipelines, routers, scoring, classifiers (@omarsar0, @Yuchenj_UW). Autoregression there is pure overhead.

Neutral. Engineers see not a replacement but a new compilation layer: expensive LLM calls get decomposed into many small typed functions in the spirit of DSPy signatures, each executed by a cheap specialised model (@eggie5, @dbreunig).

Against. Sceptics note that Jev is not a general-purpose language model: it produces no free-form text and requires a predefined output format, which makes the comparison with LLMs wrong by construction (@scaling01).

Tools and techniques

In brief