Open weights become the lock-in hedge
The fallout from Washington forcing Anthropic's Fable 5 and Mythos offline is reshaping how labs pitch AI: Sakana's Fugu and Z.ai's GLM-5.2 both lean hard on "no single vendor can switch us off" as a selling point. Meanwhile open-weights releases keep narrowing the gap with frontier closed models, and OpenAI lands one of its largest enterprise deals yet at Samsung.
Trump administration forces Anthropic to pull Fable 5 and Mythos offline
An export control order citing unspecified national security concerns required Anthropic to ensure its two newest models couldn't be accessed by foreign nationals, so the company pulled Fable 5 and Mythos entirely. Reporting ties the order to Amazon researchers who allegedly bypassed Fable 5's guardrails, with Andy Jassy raising it to the White House. Cybersecurity experts signed an open letter calling the order dangerous, arguing it strips network defenders of capabilities and that the same jailbreaks exist in other models.
Why it matters: If a frontier model can vanish overnight on a Friday-afternoon order, anyone building critical infrastructure on a single closed API now has a concrete regulatory risk to price in.
Sakana's Fugu orchestrates a swappable LLM pool to rival Anthropic's top models
Tokyo-based Sakana AI launched Fugu, a language model trained to call other LLMs from a swappable agent pool while presenting a single OpenAI-compatible API. Sakana says Fugu Ultra matches Fable 5 and Mythos Preview across coding, reasoning, science and agent benchmarks despite neither being in its pool. The company explicitly pitches the design as a hedge against vendor lock-in, citing the Anthropic export controls, though it doesn't address the token-cost overhead of orchestration.
Why it matters: Orchestration-as-a-model is a real architectural bet, but "resilience" isn't sovereignty: if several top providers restrict access at once, Fugu's options shrink with them.
- Sakana AI's Fugu orchestrates multiple LLMs to match Anthropic's Fable and Mythos benchmarks (The Decoder)
- Sakana Fugu (Hacker News)
GLM-5.2 leads open weights but loses the head-to-head to Opus 4.8
Z.ai's MIT-licensed GLM-5.2 ships with a 1M-token context and High/Max thinking tiers, and ArtificialAnalysis ranks it the top open-weights model on its Intelligence Index (51) — at roughly a fifth of Opus's output price. In a one-shot raw-WebGL 3D platformer test, Opus 4.8 was faster and shipped a cleaner, correct game; the text-only GLM-5.2 ran longer, cost far less, and shipped fundamentals broken (gray untextured character, non-lethal hazard, no win condition). Being multimodal let Opus screenshot and self-correct; GLM fell back to sampling pixel colors and missed its own bugs.
Why it matters: GLM-5.2 is the rare frontier-adjacent model no vendor can revoke, but text-only self-verification is a hard ceiling on visual tasks — and it burns ~43k output tokens per task.
- GLM 5.2 vs. Opus (Hacker News)
Swiss AI Initiative ships Apertus, a fully open foundation model for sovereign AI
EPFL, ETH Zurich and CSCS released Apertus with open weights, open data, and open training code, claiming to be competitive with top open models at 8B and 70B scale and trained on 1000+ languages. The release includes Apertus Mini, a set of 16 small models demonstrating distillation and quantization. It's positioned for EU AI Act compliance, respecting opt-outs, removing PII, and limiting memorization.
Why it matters: Reproducible open data and methods — not just open weights — is what auditors and EU-regulated deployments actually need, and it's still rare at this scale.
- Apertus – Open Foundation Model for Sovereign AI (Hacker News)
Samsung deploys ChatGPT Enterprise and Codex to all Korean staff in one of OpenAI's biggest deals
Samsung Electronics is rolling out ChatGPT Enterprise and Codex to all employees in South Korea and its worldwide Device eXperience division, which OpenAI calls one of its largest enterprise deals. OpenAI says Codex now has more than five million weekly users, with Korean active users up roughly 800% since February, and notes non-developers increasingly use it to build internal tools via a new record-and-replay feature. Samsung also supplies OpenAI with memory chips for AI infrastructure.
Why it matters: Codex is quietly becoming a general workflow-automation tool, not just a coding assistant — and the chips-for-seats reciprocity shows how entangled the supplier and customer relationships are getting.
Berkeley study: ChatGPT inflated grades in writing- and coding-heavy courses
Analyzing 500,000+ grades across 319 courses at a large public research university, Igor Chirikov found the share of A's jumped 13 percentage points after ChatGPT's late-2022 launch, concentrated in writing- and coding-heavy courses. The effect clusters in homework rather than proctored exams — courses where homework carries above-median weight saw an extra 16-point A increase — and a placebo test on oral presentations showed no movement. The author argues this reflects outsourced work, not learning gains, and warns of a feedback loop weakening graduates in exactly the skills AI is strongest at.
Why it matters: If credentials in coding-heavy programs increasingly certify AI output rather than skill, the hiring signal degrades right as AI also makes interviews easier to game.
sqlite-utils 4.0rc1 adds migrations and nested transactions
Simon Willison released the first release candidate for sqlite-utils v4, folding the proven sqlite-migrate package in directly as a built-in migrations system driven by decorated Python functions and a new migrate CLI command. It also adds db.atomic() for nested transactions backed by SQLite savepoints, borrowing Django/Peewee terminology. The major bump carries breaking changes: type detection now defaults on for CSV/TSV import, REAL replaces FLOAT, schemas use double-quotes, and db.table() no longer returns views.
Why it matters: A widely used building block for LLM data pipelines gets first-class migrations and transactions — worth testing the breaking changes before the stable release lands.
- sqlite-utils 4.0rc1 adds migrations and nested transactions (Simon Willison)
- sqlite-utils 4.0rc1 (Simon Willison)
Fine-tuning Qwen 3 0.6B turns a tiny model into a 92%-accurate classifier
A developer building a household RAG chatbot fine-tuned Qwen 3 0.6B with Unsloth and QLoRA to categorize incoming questions and narrow the vector search space. Prompting the base model alone scored just 10% on a 131-test battery; fine-tuning lifted it to 79%. Mapping categories to two-character opaque IDs with no semantic overlap — instead of human-readable labels — pushed accuracy to ~92% by eliminating fragment and confusion errors.
Why it matters: A concrete reminder that a 600M-parameter local model can handle narrow classification reliably after fine-tuning, and that output-format design often beats prompt-tweaking.