AI Daily · 2026-09-21

Hugging Face's tokenizers v1 RC is the main release, making encoding/decoding much faster while guaranteeing identical token IDs and keeping the same API. The other through-line is TypeSafe AI's non-generative Jev model, which LangChain integrated and benchmarked as an evaluator; it showed dramatically lower cost, latency and variance than LLM judges. OpenAI's V7 case uses GPT-5.6 and GPT-6 Astra to build a Context Graph that gives agents institutional memory with auditable decisions. Research on block-removal pruning framed as an Ising optimization problem reports strong MMLU gains for Llama-3.3-70B at 50% compression. Simon Willison released llm-keys-ui 0.1 to manage API keys on remote machines.

North America · First-hand

OpenAI

⭐⭐ [Other] How V7 gives AI agents institutional memory

OpenAI published a customer story describing how the European fintech startup V7 gives AI agents enterprise business context. V7 Go, V7's agentic platform, connects to repositories such as SharePoint and Google Drive, extracts entities, relationships, facts, attributes and metrics from millions of files, and organizes them into a Context Graph that links entities, relationships and cited evidence to power MCP search and workflows spanning hundreds of steps. V7 Go uses GPT-5.6 Luna for extraction, GPT-5.6 Terra and Sol for reasoning and tool use across complex multi-step instructions, and has begun using GPT-6 Astra for the most demanding Context Graph queries, including financial analysis across thousands of documents. Reported figures include 89% accuracy for GPT-6 Astra on the hardest queries, 78% lower cost per document and +11.6 points higher accuracy with GPT-5.6 Luna, and, on the HERB benchmark, a 69% improvement over the official baseline with 38% fewer hallucinations on unanswerable queries. V7 says agents complete 50–100 step workflows in minutes at 99.9% accuracy while keeping an auditable trail of every decision.

Ecosystem & Beyond (Products / Agents / Tools / Opinions)

Model Release

⭐⭐ [Model Release] What Is Jev? A Guide to TypeSafe AI’s System One Model

Jev is a new model from TypeSafe AI that is not a traditional LLM and does not generate text; the company reports up to 200x faster inference and 400x lower cost than comparable LLMs on classification tasks. TypeSafe AI calls it a "System One" model, built to make fast, structured decisions software can use directly and trained with reinforcement learning for calibrated decisions (RLCD). Users send it a state (context) and questions; it supports three question types—Choice (pick from options, returning probabilities and confidence), Score (rate against ordered levels, returning a continuous score, distribution, and confidence), and Noul (yes/no, returning the probability a statement is true)—and multiple questions in one request are evaluated in parallel with little added latency. LangChain integrates Jev through TypeSafeClassifier, where .invoke() returns classification results rather than a chat response, and the state can be text, structured data, or LangChain messages.

Product Update

⭐⭐⭐ [Product Update] tokenizers v1: encode, decode and scaling, measured

Hugging Face announced the release candidate of tokenizers v1, with the main focus on performance: the post states v1 is faster than v0.23, in some cases by tens of times, while producing the same token IDs and preserving the API, vocabulary and merge ranks. The library has been split into a workspace, where tk-encode is the required runtime and tk-serialize, tk-convert and tk-train are linked only when needed. The article presents benchmark results against other widely used alternatives, covering single-threaded and multi-threaded runs, scaling across threads, per-model and per-language comparisons, latency, decoding throughput, memory heap and crate size, with the benchmarks runnable on your own hardware from the tokbench repository. The authors note that ideas were drawn from open source projects such as gigatoken, tiktoken, kitoken, tokie, fastokens, wordchipper and ai-tokenizer, and thank IBM, NVIDIA and the ExecuTorch team for patches and hardware testing.

⭐⭐ [Product Update] llm-keys-ui 0.1

Simon Willison released llm-keys-ui 0.1, a plugin that provides an LLM web UI for setting API keys. It solves a specific problem: he runs coding agents on various machines via Codex Remote and controls them from his phone, and occasionally needs to configure an API key, but does not want to paste API keys into agent sessions or into the ChatGPT app. The workflow is to have Codex run `uvx --with llm-keys-ui llm keys-ui --all`, which reports a URL — including local network or Tailscale device IPs — for an interface where additional API keys can be saved. Keys can later be pulled into shell commands with something like `llm keys get anthropic`.

Research

⭐⭐ [Research] Pruning LLMs Like a Physicist: Block Removal as an Ising Optimization Problem

The post presents the paper 'LLM Compression by Block Removal with Constrained Binary Optimization' from Multiverse Computing. Block removal (depth pruning) deletes whole transformer blocks to get predictable inference speedups that stack with quantization and low-rank compression, but the hard part is choosing which blocks to cut: the effect of removing one block depends on which others are removed alongside it, making it a combinatorial rather than a ranking problem. The authors reformulate block selection as a constrained binary optimization that maps onto an Ising glass with all-to-all interactions and a fixed number of 'up' spins; the energy of that spin system is a cheap proxy for the pruned model's benchmark score, so many candidate configurations can be ranked without benchmarking any of them. At 50% compression of Llama-3.3-70B-Instruct, the method gains almost 23 percentage points on MMLU over the best competing block-removal method. Each block is assigned a binary variable (0 keep, 1 remove), and a second-order Taylor expansion of the model loss yields an approximate Hessian whose diagonal gives per-block importance and whose off-diagonal entries give the pairwise couplings between blocks.

⭐⭐ [Research] Can Jev Be a Better Agent Evaluator?

LangChain published a blog post testing whether TypeSafe AI's new model, Jev, can serve as an agent evaluator compared with LLM-as-a-judge. Jev is not a traditional autoregressive LLM and does not generate text; it is what TypeSafe AI calls a "System One" model that evaluates a state and returns typed answers with probabilities. In LangChain's experiment, Jev was far more consistent on continuous scoring, with 92–913x lower quality-score variance than GPT-5.6 Luna, Terra, and Claude Sonnet 4.6, and it was also the fastest and cheapest at an average 0.44s and $0.00035 per call ($0.34 total versus $28.17 for Claude). According to TypeSafe AI, Jev is up to 200x faster and 400x cheaper than comparable LLMs on classification tasks. LangChain says the results are promising but early, coming from a narrow test, and points to a new direction for agent evals; the post also contrasts the limitations of code-based evaluation and LLM judges.