Open Source Releases

hf-xet 1.5.3.dev0: Fast Large File Transfer for Hugging Face Hub — Implements the XET protocol for accelerated transfers of multi-GB model weights and datasets on HF Hub. Content-addressable storage plus parallel chunked transfers means less time staring at progress bars. 🤖

lex-oracle - Deterministic Legal Computation for Slovak/Czech Law — MCP server that serves hallucination-free legislative blueprints for payroll, tax, garnishment, travel, and B2B computations. Uses citable deterministic sources instead of LLM generation — because nobody wants their tax code hallucinated. 📄

anthropics/claude-code v2.1.218 — /code-review now runs as a background subagent so review work stops cluttering your main conversation. Also added screen-reader announcements for deleted text across a bunch of keybindings. Accessibility wins. 🛠️

block/goose v1.44.0 — Latest Gemini models supported, plus a declarative Sakana AI provider for OpenAI-compatible Fugu API. Stop hook context now receives working_dir for cleaner session cleanup. Incremental but useful. 🤖

sst/opencode v1.18.4 — Adaptive thinking controls for Kimi models on Anthropic-compatible providers with summarized reasoning by default. Fixed OpenAI provider header timeouts during slow connection setup. The timeout fix alone might be worth the upgrade. 🛠️

glia-agents 0.8.1: Minimal Glass-Box LLM Agent Library — Transparent, minimal library exposing modern agent techniques as opt-in primitives. No hidden control flow — you see exactly what executes. For engineers who’ve been burned by opaque agent frameworks. 🔍

Research Worth Reading

JAXBench: Benchmarking Autonomous TPU Kernel Optimization — 50 JAX workloads as a TPU-native benchmark for AI-generated kernel optimization on Google Cloud TPUs. Finally a standardized target for TPU kernel work — GPU benchmarks have had this for ages. 📄

SonicSampler: Unified Tile-Aware Kernels for LLM Sampling and Speculative Verification — Single kernel fusing logit processing, token selection, and speculative verification. Eliminates multiple kernel launches and handles dynamic sampling behavior across batches. Kernel fusion done right. 🔥

DC-Leap: Training-Free Acceleration of dLLMs via Draft-Guided Contiguous Leaping Decoding — Training-free decoding speedup for diffusion LLMs that sidesteps conservative confidence thresholds from Joint Probability Dependence Error. Draft-guided leaping cuts redundant denoising iterations. Free speed is the best speed. 📄

InferenceBench: A Benchmark for Open-Ended LLM Inference Optimization by AI Agents — Evaluates AI agents on open-ended inference optimization, not prescribed workflows. Finally a benchmark that can’t be solved by memorizing known recipes. About time. 📄

PhantomFill: When the Form Demands an Answer, Language Models Invent One — Structured output formats (JSON, function args, extraction templates) themselves induce hallucination. Tested 13 models on unanswerable questions — format constraints fabricate responses. Your JSON schema is lying to you. 🔥

Robust Critics: Defending LLMs Against Multi-Turn Attacks — Critic-based defense that distinguishes genuine attacks from misunderstood queries across multi-turn dialogues. Maintains safety without nuking legitimate users. The compounding ambiguity problem is real. 📄

AI Dev Tools

Alibaba Open Code Review: Hybrid LLM + Deterministic Code Review Tool — Battle-tested at Alibaba scale. Combines deterministic analysis pipelines with LLM agents for precise line-level comments. Built-in fine-tuned ruleset catches NPE, thread-safety, XSS, SQLi. Works with OpenAI and Anthropic APIs. 🛠️

Today’s Synthesis

PhantomFill’s finding that JSON schemas and function-calling formats actively induce hallucination should change how you architect LLM pipelines. The fix isn’t better prompting — it’s deterministic verification layers. lex-oracle demonstrates this for legal computation: citable, deterministic sources replace LLM generation entirely for payroll and tax logic. Alibaba’s Open Code Review takes the hybrid approach to production at scale, running deterministic analysis pipelines (NPE, thread-safety, XSS, SQLi detection) alongside LLM agents for line-level comments. The pattern is clear: use LLMs for what they’re good at (reasoning, synthesis, code generation) and deterministic tooling for what they’re not (arithmetic, constraint validation, security rules). Next time you reach for response_format: {type: "json_object"}, ask whether a post-generation validator — schema validation, type checking, or a rule engine — would catch the fabrications PhantomFill proves will appear. Your JSON schema isn’t a contract; it’s a hallucination trigger. Build the deterministic guardrail instead.