Automated AI News Brief: Lean Community Certifies Riemann Hypothesis Proof, Transformers v5.24.0 Unifies Cache API, and Local Quantization Trade-Offs
November 22 AI news brief: Lean and Mathlib communities certify the formalized Riemann Hypothesis proof across independent kernels, Hugging Face Transformers v5.24.0 adds 10 models and unifies the cache API, LookAtMyQuant squeezes Inkling-2-Mini into 5.4 GB, plus GPGME security patching and developer tooling updates.
Introduction
Horizon gathered the material for this post, and Codex selected and rewrote it; Horizon is used only for data collection. Today's briefing focuses on two converging fronts in AI engineering: the push for independent, machine-checkable verification as AI-assisted formal proofs reach unprecedented scale, and rapid progress in open-weights multimodal models, quantization compilers, and core developer infrastructure.
Lean and Mathlib Communities Certify Formalized Riemann Hypothesis Proof Across Independent Kernels
Following Anthropic researcher Levent Alpöge's Claude-assisted Riemann Hypothesis announcement, the Lean and Mathlib communities issued a joint statement—endorsed by the Lean FRO and the Mathlib Initiative—certifying both the theorem statement and the correctness of Alpöge's roughly 17-million-line Lean proof. Using an isolated file importing only Mathlib, the community confirmed with two independent tools (comparator and con-leche --compare) that the proved theorem matches Mathlib's vetted RiemannHypothesis definition exactly.
MIT professor Andrew Sutherland also published full verification logs generated with the help of Claude Code: all 175,187 modules compiled cleanly on Lean v4.33.1, the headline theorems rely solely on Lean's three standard axioms (propext, Classical.choice, Quot.sound), and the exported proof passed Lean's official kernel, lean4lean, the independent Rust kernel nanoda (finishing in 27 minutes 42 seconds on 80 threads), and the formally verified checkers con-leche and con-ron. Latent Space's AINews highlighted the striking contrast between Alpöge's 1.3 GB machine-verified proof and Alessandro Masullo's 940-page human draft written over 12 years—underscoring an emerging divide between proofs that are formally sound yet unreadable by humans, and human proofs that are understandable yet extraordinarily hard to audit. Meanwhile, Hacker News commenters cautioned against over-relying on a single AI ecosystem across proofs, scripts, and tooling, emphasizing the value of diverse, independently authored proof checkers.
Source: Lean Community Blog: Lean verification of the Riemann Hypothesis Source: MIT Math (Andrew Sutherland): Verification files for the Lean proof of the Riemann Hypothesis Source: Latent Space: AINews Riemann Hypothesis proved...twice independently
Hugging Face Transformers v5.24.0 Adds 10 New Architectures and Unifies the cache API
huggingface/transformers shipped v5.24.0, introducing 10 new model architectures alongside major runtime and API cleanups. New additions span looped reasoning, multimodal, and multilingual speech models: Thinking Machines Lab's Inkling 2 Mini (24 decoder layers with layers 6–17 looping three times and 8-of-128 MoE routing), Swiss AI Initiative's Apertus 1.5 (combining EMU3.5 vision tokenization and WavTokenizer for text, image, and audio inputs), NVIDIA's latent-recurrent Dory, Tencent's HY4-VL (Hunyuan V4 vision-language with dynamic-resolution mRoPE), Meta FAIR's Omnilingual ASR covering over 1,600 languages (300M to 7B in CTC and Llama-decoder variants), Translated's hybrid MoE Lara Mini, Sarvam AI's SarvamMLA (105B) and SarvamMoE (30B), and native support for LG AI Research's EXAONE models without requiring trust_remote_code=True.
On the breaking-changes front, past_key_values and cache_params are now unified under a single cache argument across model forward signatures, generate, ModelOutput, and masking utilities, bumping the minimum peft requirement to >=0.22.0 (with temporary backward-compatibility warnings). Generative multimodal models now return shared output classes such as MultimodalCausalLMOutputWithPast, and the deprecated modeling_attn_mask_utils.py file has been removed. The release also fixes continuous batching under torch.inference_mode(), resolves TextStreamer lagging on spaceless languages such as Chinese, Japanese, and Korean, adds a python-call-list chat content parser, and hardens exporter security by requiring trust_remote_code=True when loading pickled .pt2 programs.
Source: GitHub: huggingface/transformers v5.24.0
LookAtMyQuant Compresses Inkling-2-Mini to 5.4 GB Memory Amid Speed and Benchmark Debates
Open-source quantization compiler LookAtMyQuant drew strong Hacker News interest after shrinking Thinking Machines' Inkling-2-Mini (a 26.4B total / 5.4B active 3×-looped MoE model) from 18.89 GB of text weights down to four Apple Silicon MLX builds starting at 5.38 GB (1.65 bpp), alongside 8.9 GB, 11.2 GB, and 12.8 GB tiers. By combining 4 KB SRAM hashed block-vector trellis codes, Fisher-information per-row bit allocation, SQYFR proximal fine-tuning, and lossless NVFP4 passthrough on the most sensitive expert buckets in larger tiers, the project reports near-teacher GSM8K and HumanEval scores on its 11.2 GB build.
However, developer discussion in the HN thread offered important reality checks for local AI practitioners. Commenters warned against taking third-party quantization claims—particularly narrow KL-divergence metrics or small benchmark subsets—at face value without checking real-world generalization. Early testers running Inkling-2-Mini on workstation GPUs (such as the R9700) also noted noticeably sluggish throughput—roughly one-third the speed of Qwen 27B—likely due to repeated looped-transformer passes or unoptimized kernels, highlighting the ongoing trade-off between memory footprint and inference latency.
Source: GitHub: ShinMegamiBoson/lookatmyquant
Developer Tooling & Language Updates: gopgext, Zero C#, and Immutability in the LLM Era
Several developer tooling and programming language projects also stood out on Hacker News. gopgext is an experimental compiler toolchain built on TinyGo that lets engineers write PostgreSQL extensions in plain Go: annotating a function with //gopgext:function automatically generates the C glue, install script, and control files, with support for SPI SQL queries, triggers, GiST/GIN/BRIN index access methods, and ports of pgcrypto and pg_stat_statements. While the author stresses it is not yet production-ready—and developers noted Rust's pgrx remains the battle-tested choice—it lowers the barrier for Go teams extending Postgres.
Meanwhile, Show HN: Zero C# forks bflat to compile modern C# (including C# 15 unions, generics, and records) into tiny native binaries (~11 KB on Windows, ~40 KB on Linux, plus WebAssembly and 386 DOS) with no .NET runtime or garbage collector, replacing GC with explicit arena allocation. Rounding out the language discussion, a resurfaced 2017 essay on how Clojure reshapes programming intuition sparked fresh LLM-era commentary: developers observed that Clojure-style immutable data structures and explicit data transformations simplify state and shrink context requirements, making codebases easier for both human engineers and coding agents to reason about safely.
Source: gopgext: PostgreSQL Extensions in Go Source: Zero C#: C# without a runtime or GC Source: Eli Bendersky: Clojure will affect the way you think about programming (2017)
GnuPG Patches Critical GPGME Double-Free Vulnerability (CVE-2026-129975)
The GnuPG project released GPGME 2.3.1—alongside updated Gpg4win, GnuPG Desktop, and GnuPG VS-Desktop packages—to patch a critical double-free vulnerability (CVE-2026-129975, tracked as T8557) affecting all versions prior to 2.3.0. Discovered by Johan Carlsson and reported through GitLab, the bug stems from two internal buffer-reading functions for gpg status lines and key listings: when reallocating past the initial 1024-byte buffer, an early error return leaves the context object holding a stale pointer to freed memory.
A crafted OpenPGP key with a very long user ID can trigger the double free to exfiltrate application data, and GnuPG maintainers note that remote code execution (RCE) cannot be ruled out. Because GPGME is widely linked across email clients, Samba, Flatpak, and signed-commit verification pipelines such as GitLab—which accelerated disclosure alongside a Ruby2Shell update—engineering and DevSecOps teams should patch affected environments immediately.
Source: GnuPG Announce: GnuPG Security Advisory for GPGME (T8557, CVE-2026-129975)
Today's Takeaway
As AI systems begin producing 17-million-line mathematical proofs and high-volume code changes that exceed human line-by-line review capacity, deterministic verification layers independent of the model—whether standalone proof kernels, immutable data flows, or memory-safe core tooling—are becoming indispensable. At the same time, unified inference APIs like transformers v5.24.0 and aggressive local quantization tools continue expanding where models can run, reminding builders to weigh latency, generalization, and supply-chain security alongside raw capability.

