GLM 5.2 is free, frontier, and filling the Fable void

June 21, 2026

21 topics · 14 sources

Industry
The Rundown AI Daily Brief

Fable 5 Ban: US Export Control Standoff Enters Week Two

Commerce Secretary Howard Lutnick ordered Anthropic to stop distributing Fable 5 and Mythos 5 to "foreign persons" after reports of access reaching a South Korean firm with suspected China ties.[1]The Rundown, Jun 18 Markets expected a quick Monday resolution; instead the lockout lasted the full week with only positive signals by Friday.[2]AI Daily Brief Over 100 security experts signed an open letter arguing the ban "backfires" on US security interests, while Dario Amodei attended G7 meetings in France to discuss AI regulation.[1]The Rundown, Jun 18

Read more

The Order

The trigger was a Washington Post report that Fable 5 and Mythos 5 access had expanded to a South Korean firm suspected of having China ties. Lutnick's letter warned Anthropic against distributing the models to "foreign persons" — a broad definition that effectively required pulling global access. The White House also demanded Anthropic block all jailbreaks, which security experts called technically infeasible.

Internal Reaction

Internal Anthropic messages showed employees feeling "unfairly targeted" and "bullied based on bad vibes." The Rundown framed it as "a relationship issue more than a safety one." G7 European leaders sought access to Fable/Mythos during the standoff while simultaneously planning AI sovereignty strategies — a pointed irony the AI Daily Brief highlighted.

"For the first time in around 3 years it feels like the AI table has been flipped over." — Mike McNally, USV

Market Effect

The ban immediately accelerated trends that were already building: Chinese open-weight models (primarily GLM 5.2) filled the capability gap, OpenRouter launched its Fusion API as a multi-model routing alternative at half the cost of Fable, and developers began treating single-model dependency as a new category of risk. As of June 19, access had not been restored.

Tools: Fable 5, Mythos 5, OpenRouter Fusion
AI Models
Developers Digest AI Search AI Daily Brief The Rundown

GLM 5.2 Passes the Frontier Vibe Check

Z.AI's GLM 5.2 — a 753B-parameter MoE model released under the MIT license — scored near GPT-5.5 and Fable 5 on the Artificial Analysis Intelligence Index (~51) while outperforming both on VendingBench, a long-horizon business simulation benchmark.[3]Developers Digest Jeremy Howard called it "a marvel — at least as good as Opus 4.8 and GPT 5.5. Super fast, inexpensive, and not too verbose. I've never experienced an open-weights model like this before."[2]AI Daily Brief Its release timing — the same week Fable went dark — turned it into the community's default Fable alternative.[1]The Rundown, Jun 17

Read more

Model Specs

GLM 5.2 is a mixture-of-experts architecture with 753B total parameters (40B active) and a 1M token context window. Full weights are 1.51 TB under MIT license. Within two days of release, Unsloth published compressed GGUF variants: a 1-bit version (223 GB, 76% accuracy) and a 2-bit version (245 GB, 82% accuracy), enabling consumer-hardware deployment.[4]AI Search

Benchmarks

On the Artificial Analysis "Briefcase" long-horizon knowledge benchmark (released after GLM 5.2, preventing gaming), it ranked third — behind only the best Claude model and above top GPT models. On VendingBench (simulated vending machine business, year-long horizon), it beat both GPT-5.5 and Fable 5. Hallucination rate is 50% lower than Fable and 3–4x lower than GPT-5.5 on Artificial Analysis tests. It ranked 10th on LM Arena due to lower steerability, and performs less well on LiveBench reasoning. On DeepSqueeze (coding agent tasks), it trails Fable 5, GPT-5.5, and Opus 4.8 but at roughly half the cost.[3]Developers Digest

Pricing and Access

Input: ~$1.40/M tokens; output: ~$4.40/M tokens — roughly 42 cents per intelligence task versus 83 cents for GPT-5.5. Available via Hugging Face, OpenRouter, OpenCode, Fireworks, Factory, and Devin. OpenCode subscription is $5/month first month with ~$60 of inference included.

"This is the first model that passes the vibe check. This feels like a DeepSeek R1 moment that will push the frontier labs into releasing even better models." — Riley Brown
"You can't bench max for a benchmark that didn't even exist before." — AI Search
Tools: GLM 5.2, OpenRouter, OpenCode, Unsloth, Fireworks, Factory, Devin
AI Models
The Batch #358

Fable 5's Benchmark Numbers Need an Asterisk

Fable 5 routes flagged questions about cybersecurity, biology, and AI safety to a Claude Opus 4.8 fallback — meaning published benchmark scores include Opus 4.8 answers, not Fable 5's alone.[5]DeepLearning.ai The Batch #358 On graduate-level science (GPQA), the difference is stark: 93.18% with fallbacks included versus 55.56% on refusal-heavy topics without.[5]DeepLearning.ai The Batch #358 Evaluators also note that Anthropic's 30-day data retention policy discourages proprietary benchmark testing.

Read more

The Batch's reporting reveals that Fable 5's safety classifier intercepts queries it deems sensitive and silently routes them to Claude Opus 4.8. This means evaluation results vary dramatically depending on how evaluators count refusals and which prompt phrasing triggers the fallback. For researchers trying to compare Fable 5 fairly against other frontier models, the hidden routing creates a fundamental measurement problem — the benchmark score belongs to a system, not a model.

The 30-day data retention policy is a separate friction point: evaluators running proprietary test suites can't store results beyond a month, making longitudinal comparisons hard and deterring independent academic benchmarking.

Industry
The Rundown AI Daily Brief

SpaceX Acquires Cursor for $60B

SpaceX completed the acquisition of AI coding IDE Cursor for $60 billion in an all-stock deal, leveraging the valuation lift from SpaceX's IPO earlier in the week.[1]The Rundown, Jun 17 Cursor CEO Michael Truell indicated their next model will be "generally intelligent" at a scale comparable to Claude Opus — suggesting a 700B+ parameter model in development.[2]AI Daily Brief The deal puts one of the leading AI coding tools under Elon Musk's umbrella.

Read more

The SpaceX IPO last Friday nearly doubled the company's valuation, giving Elon the leverage to close the Cursor deal at a $60B all-stock price. Truell's hint about their forthcoming model — "generally intelligent" and "Opus-scale" — aligns with reports that the model is projected around 750B+ parameters, which would put Cursor's own frontier model in the GLM 5.2 / Opus 4.8 tier. The acquisition raises questions about pricing strategy: whether Cursor will continue competing on cost and performance, or leverage the SpaceX/xAI relationship to access proprietary compute at scale.

Developer Tools
Better Stack

TypeScript 7 RC: Compiler Rewritten in Go for 10x Speed

TypeScript shipped a release candidate for version 7, rewriting the compiler from TypeScript to Go and achieving roughly 10x faster type checking — a benchmark on Playwright's repo dropped from 6 seconds to 0.87 seconds across 1,400 files and 500K lines.[6]Better Stack The Go rewrite enables shared-memory parallelism; new flags (--checkers, --builders) control how many parallel type-checking and project-reference build workers are used. Breaking changes from TS6 → TS7 are minimal — most were introduced in TS5 → TS6.

Read more

What Changed

The old JavaScript compiler was single-threaded; Go's shared-memory model enables parallel type-checking workers (default 4 via --checkers) and parallel project reference builds (--builders). Watch mode was also rewritten using a port of Parcel's C++ file watcher. A compatibility package lets TypeScript 6 and 7 coexist in the same project, and a stable programmatic API for tooling authors lands in 7.1 rather than 7.0.

Breaking Changes

The only new language change in 7.0 itself: template literal types now split on Unicode code points rather than UTF-16 code units (fixing a long-standing correctness bug). All the heavier breaking changes happened in TS5 → TS6: ES5 target removed, AMD/UMD/SystemJS dropped, strict: true as default, ES Next as default module. Going from 6 to 7 should be low-friction for most projects.

Stable release expected within a month. The programmatic API (used by Prettier, ESLint, and other tooling) is held for 7.1 to give the ecosystem time to adapt.

Tools: TypeScript 7, Go, Bun, Playwright
Podcast
Lenny's Podcast

Lenny's Podcast: Fiona Fung on 8x Code and What Comes After

Fiona Fung, who leads Claude Code and Cowork at Anthropic, joins Lenny Rachitsky to explain what it actually looks like to run a product team where coding is no longer the bottleneck — Anthropic engineers now ship roughly 8x as much code per quarter as in 2021–2025, and PMs and designers check in code too.[7]Lenny's Podcast With throughput this high, verification and staying on top of what is actually shipping become the new constraints. She also explains what Cowork is — Anthropic's product extending Claude's agentic capabilities to non-coders — and how a personal frustration with expense reports turned into a small-business product launch.

Read more

Sections

~02:02 Fiona's 25-year path: IBM, Visual Studio, Meta, dog-fooding
~08:06 The AI-pilled team: 8x code, everyone ships, verification as the new bottleneck
~10:06 Remote Claude sessions and routines to stay on top of all the shipping
~14:09 Automated code review and specs-in-repo as the evolution of TDD
~26:13 Closing the AI divide: Cowork, small businesses, and latent demand
~35:16 The next frontier: async work, fleets of agents, routines spawning agents
~39:18 High agency + high accountability; measuring ROI beyond tokens and lines of code
~44:23 Proactive quality: "bad vs sad" framework, monitoring, evals, swear-word dashboard
~55:32 What's lost: flow, loneliness, growing the next generation
~78:49 Maintaining culture through explosive growth

How She Stays on Top of All the Shipping

Fiona runs a persistent remote Claude Code session enrolled in all repos and connected to Slack channels and metrics dashboards. It serves two purposes: generating PRs and bug fixes, and powering monthly reviews where she has data-grounded conversations with the team about what shipped, how it performed, and where quality hot spots are forming. She also uses "routines" — scheduled processes that monitor feedback channels, generate prompts, spawn agents, and leave reviewable PRs waiting when she wakes up. This async agent fleet model is where she sees engineering heading.

Quality at Scale

Quality is managed proactively through specs checked into the repo ("what good looks like" — she calls this an evolution of TDD, where the model now writes the test), automated Claude code reviews against those specs, monitoring, and evals. The "bad vs sad" error taxonomy lets teams set their own quality thresholds with high agency. The team has a "swear-word dashboard" that catches crude or inappropriate error messages early.

Cowork and the AI Divide

Cowork extends Claude's agentic capabilities to non-coders for general knowledge work. Fiona's personal trigger use case was expense reports; she onboarded small-business-owner friends (a restaurateur) who turned a messy downloads junk drawer into searchable menus and ran competitive price analysis — which surfaced latent demand and fed into a "Claude for small business" toggle inside Cowork.

"Coding is no longer the bottleneck... it's lifted the ceiling of what anyone is able to do."
"Don't forsake motion for progress. If you're measuring tool usage, then you're measuring the action, but is it really making the end outcome important?"
"With high agency is also high accountability. It's all about making sure folks have that freedom to cook, but then it's also, okay, what's the accountability for it?"
Tools: Claude Code, Cowork, Claude for small business, Routines, Claude Code Reviews, Slack
Podcast
Latent Space

Latent Space: Ronak Malde on Living Systems and Trajectory.ai

Ronak Malde (ex-Windsurf/Codeium, ex-DeepMind) left $2B+ acquisition money on the table to found Trajectory.ai on the thesis that every future product will be a "living system" — one that continually learns from real-world user behavior rather than staying static like conventional software.[8]Latent Space Trajectory is a self-serve continual-learning platform that ingests expert agent traces, generates evals and training environments, and retrains a company's model end-to-end — with Harvey and Nvidia Nemotron-3 Super as an early production example. He also introduces SDPO (Self-Distillation Policy Optimization) and an open-source continual LoRA training stack built with Berkeley's SkyRL lab.

Read more

Sections

~02:06 Building SWE-1 and the foundation-model flywheel at Windsurf
~03:30 The DeepMind acquisition: Demis and Sergey on the call, 35-day turnaround
~06:10 Why leave DeepMind: static AI vs. continual learning thesis
~11:13 Trajectory platform, Harvey + Nvidia Nemotron-3 Super
~15:16 Data curation: directional corrections vs. noisy binary signals
~17:19 SDPO: self-distillation policy optimization explained
~22:28 Continual LoRA: open-source concurrent training stack with SkyRL/AnyScale
~27:32 Act two: legibility, control, and scaling to Fortune 500

The Core Thesis

The model you used yesterday makes the same mistakes tomorrow — all user corrections are wasted. Products like SWE-1 at Windsurf, Cursor Composer, and Claude Code pull ahead because they build the model around the product and compound from real-world usage. Trajectory wants to give every company that same loop.

The Platform and Harvey

Trajectory ingests expert agent traces and distills them into a unified "trajectory" format, then generates evals, judges, and training environments. Key data curation insight: binary signals (thumbs up/down, accept/reject) are noisy because users vibe-accept everything; the signal is in directional corrections (moving a button, redlining a contract). For Harvey, Trajectory partnered with Nvidia to train Nemotron-3 Super (12B) to the Pareto frontier on legal workflows — issue spotting, citation, completeness — at a fraction of frontier model cost. "For a field like legal, getting 80% of the way there is the same thing as zero." Customer onboarding: 3 months for the first → under a week for new ones.

SDPO and Continual LoRA

SDPO (Self-Distillation Policy Optimization) avoids collapsing correction signal into a single RL reward. Instead, the teacher is the same model given privileged hints from production; the student learns to match the teacher's log-probs — learning from actual text guidance. The open-sourced continual LoRA stack (with Berkeley SkyRL / AnyScale) breaks the linear train-teardown cycle into dedicated concurrent sampling and training pools, cutting wall-clock time roughly in half.

"The cool unlock with self-distillation is you don't actually need a smarter teacher. Your student could be the smartest model, but to make the teacher smarter, I'm going to take some privileged information or a hint and put that into context of the teacher."
"No one's really made the product for when intelligence is just getting smarter every single day."
Tools: Trajectory.ai, Windsurf, SWE-1, Nvidia Nemotron-3 Super, Harvey, Clay, Rogo, SDPO, SkyRL, AnyScale
AI Tools
Nate B Jones

Nate B Jones: You Can't Run AI Agents Without an Owner

Nate B Jones argues that the 2026 AI skill isn't building agents — it's maintaining them: every agent needs a specific job (one sentence), a curated diet of context, clear boundaries on what it can touch, and a review loop to catch drift.[9]Nate B Jones The danger of unowned agents isn't dramatic failure — it's quiet, invisible drift over time (stale sources, plausible-but-wrong outputs) that compounds because nobody checks it. He closes with a six-field Owner Card template and a decision rule: if nobody is willing to own an agent, it probably shouldn't be doing important work.

Read more

Sections

~00:00 What actually makes something an agent (practical definition)
~03:00 The risk of unowned agents: silent drift and invisible consequences
~03:50 The four pillars: job, diet, boundaries, review loop
~04:50 The Pokémon analogy: collecting isn't the goal, knowing is
~06:02 Product team walkthrough: the story prep agent
~08:50 Prompts vs. jobs: the core 2026 mindset shift
~09:50 Agent roster for team leaders
~11:04 The Owner Card template

The Four Pillars

Job: must be expressible in one sentence ("draft refund replies for this ticket type" not "help with product"). Diet: what the agent reads — docs, tickets, transcripts; stale inputs produce stale outputs. Boundaries: read-only → draft-only → write to systems of record, each tier a categorically higher risk. Start narrow and let the agent earn permissions. Review loop: agent runs, a human reviews, the owner updates instructions and sources, repeat.

Prompts vs. Jobs

A prompt is "Write acceptance criteria for this feature." A job is: "Read the PRD, the last 20 support tickets, the design brief, and our three best backlog examples. Draft the stories for refinement, attach the customer evidence, mark assumptions, and don't create Jira tickets — put everything into review so I can look at it first." The job shapes how the whole team works, not just one individual's productivity in the moment. Skill progression: prompting (2023) → delegation (2025) → maintenance (2026).

Agent Roster and Owner Card

Team leaders need a simple roster — a list of every running agent with: name, owner (one person), job, sources, permissions, review cadence, known failure modes. Without it, "work moves through tools and nobody can explain how the output got there." Owner Card fields: name, owner, job (one sentence), sources, what it can do, what it cannot do, failure mode to watch for. Decision rule: if nobody is willing to own it, decommission it.

"Collecting the Pokémons isn't the point. You have to know what each Pokémon is good at, what it's been trained on, where not to use it, and when it starts picking up bad habits."
"Just building a new agent, you shouldn't get credit for these days. Owning an agent and using it to deliver value — that's what you should get credit for."
Tools: Claude Code, Codex, Claude Projects, Jira, Slack
AI Models
The Batch #358

Nvidia Nemotron 3 Ultra: 550B, 183 Tok/s, Open Weights

Nvidia released Nemotron 3 Ultra, a 550B-parameter hybrid transformer-mamba model that generates at 183 tokens per second — roughly 3x faster than comparable open-weight models — with a 1M token context window and 95% recall.[5]DeepLearning.ai The Batch #358 It scores 47.7 on the Artificial Analysis Intelligence Index and is released under the OpenMDW-1.1 license with weights and training data included. Trajectory.ai used an earlier Nemotron (Nemotron-3 Super, 12B) in their Harvey legal AI deployment.

Read more

The hybrid transformer-mamba architecture is what enables the speed advantage — mamba layers handle long-range dependencies without the quadratic attention cost, allowing generation at 183 tok/s even at 550B scale. The 1M token context with 95% recall is practical for legal, financial, and research document workflows where frontier models either lack context or cost too much at scale. OpenMDW-1.1 is a permissive license that includes both model weights and training data, making it usable for continued pretraining and research.

AI Models
The Batch #358

Three Agentic Benchmarks to Replace SWE-Bench

Three new agentic benchmarks targeting gaps in SWE-Bench's coverage: DeepSWE (multilingual feature implementation — GPT-5.5 at 70%, Opus 4.8 at 58%); ProgramBench (generate 200 programs from descriptions — best model hits only 3% at the 95% correctness threshold); ITBench-AA (infrastructure log diagnosis — Opus 4.7 at 46.7% full recall).[5]DeepLearning.ai The Batch #358 The ProgramBench number is the most striking: the best available model barely works 3% of the time at near-perfect correctness, suggesting a large gap between benchmark-saturated headline scores and actual program synthesis capability.

Read more

DeepSWE extends SWE-Bench with multilingual tasks (Python, JavaScript, TypeScript, Java) and evaluates models on implementing real GitHub issues end-to-end. GPT-5.5 leads at 70%; Opus 4.8 at 58%.

ProgramBench tests whether models can generate correct complete programs from natural language descriptions — 200 programs, scored at a 95% test-passing threshold. The 3% ceiling reveals that models good at patching code still struggle to synthesize correct programs from scratch at production quality.

ITBench-AA evaluates log-based infrastructure diagnosis: given system logs from a failing production environment, can a model identify the root cause? Opus 4.7 achieves 46.7% full recall, meaning over half of diagnosed incidents are still missed or incomplete.

Industry
OpenAI Blog

Samsung Deploys ChatGPT Enterprise and Codex to All Employees

Samsung Electronics is rolling out ChatGPT Enterprise and Codex to all employees in Korea and all global Device eXperience (DX) division staff — described by OpenAI as one of its largest enterprise launches ever.[10]OpenAI Blog Codex weekly active users in Korea grew nearly 800% since February 1, 2026; more than 5 million people use Codex every week globally. Use cases span software development, marketing, product development, and manufacturing.

Read more

The Samsung deployment covers all Korean employees (both ChatGPT Enterprise and Codex) plus all worldwide DX division staff. Samsung and OpenAI also have a separate collaboration on AI memory semiconductors for AI infrastructure. Seoul National University deployed ChatGPT Edu to all 47,000 members simultaneously. Other Korean companies now on OpenAI's enterprise roster: LG Electronics, LG Uplus, LG CNS, GS E&C, Samsung SDS, TVING, Krafton, and Toss.

Tools: ChatGPT Enterprise, Codex, ChatGPT Edu
Productivity
AI Search

OpenAI Codex Record and Replay: Teach by Showing

OpenAI's Codex now lets users screen-record any computer workflow once; Codex watches the video, infers the steps and context, and generates a reusable automation skill — no manual scripting or prompt writing required.[4]AI Search A demo showed a video publishing workflow automated end-to-end: matching spreadsheet rows, filling metadata, adding captions, uploading, and verifying — all triggered on the next video package without further input. Currently macOS-only with computer use; unavailable in the EU; Windows support planned.

Read more

Record and Replay shifts the burden of automation authoring from writing prompts and scripts to simply demonstrating the workflow once. Codex infers not just the steps but the context — where metadata lives, how files are organized, what constitutes a successful completion. The demo produced a near-zero-touch automation that handled the next video package autonomously. Best for stable, well-defined workflows with clear success criteria.

"This makes automation feel less like randomly prompting an AI agent and hoping it gets it right — more like training an AI assistant by just demonstrating it yourself."
Tools: OpenAI Codex
Industry
AI Engineer

AI Engineer World's Fair 2026: Finance Is the Next Code

Swyx previews the AI Engineer World's Fair 2026 at Moscone West — 4x bigger than 2025 with 600+ attendees, 10–12 simultaneous tracks, and 41 side events — and makes the case that finance is the AI vertical most likely to break out after code.[11]AI Engineer New additions include a "token billionaire" lounge for companies spending 1B+ tokens/month, an ACM-co-organized academic conference on AI and agentic systems, and four deployed-AI verticals tracks (healthcare, finance, GTM, commerce).

Read more

The leadership track occupies 50% of breakout space, with thematic daily focuses on token maxing and AI factories. The finance vertical is the "big bet" — Swyx plans a dedicated AI Engineer New York event focused on finance, and argues that if there's any vertical after code that's closest to breakout, it's finance. The poster session format allows Twitter hot takes to be defended alongside research papers at the academic track. July 2nd is an extra content day on top of the main conference.

"The talks are kind of optional. What you want to do is the hallway track."
"If there was a vertical after code that is the most likely and closest for takeoff, it is finance."
Developer Tools
Simon Willison

sqlite-utils 4.0rc1: Migrations and Nested Transactions

Simon Willison released sqlite-utils 4.0rc1 with two features ported from other projects: a migrations system (from sqlite-migrate) for version-controlled schema changes via decorated functions, and nested transactions via db.atomic() (borrowing Django's "atomic" terminology) that wraps SQLite savepoints with automatic rollback.[12]Simon Willison Breaking changes: upsert now uses native INSERT … ON CONFLICT SET, Python 3.8 dropped, default float type changed from FLOAT to REAL, type detection is now default for CSV/TSV imports.

Read more

The migrations system deliberately omits reverse migrations — all fixes must go forward, which Simon considers the right design for most production use cases. The system has been tested in production by the LLM tool. db.atomic() uses SQLite savepoints under the hood, meaning nested calls create savepoints rather than actual nested transactions, with automatic rollback on exception.

Full breaking changes list: (1) upsert uses INSERT … ON CONFLICT SET syntax; (2) Python 3.8 dropped, Python 3.13 added; (3) db.table() now only works with tables (use db.view() for views); (4) default float column type changed from FLOAT to REAL; (5) type detection is now the default for CSV/TSV imports; (6) identifiers wrapped in double-quotes instead of square-braces.

Tools: sqlite-utils, sqlite-migrate, LLM
Developer Tools
Simon Willison

Cloudflare Temporary Accounts: 60-Minute Ephemeral Deployments

Cloudflare launched temporary accounts that let developers deploy Workers projects with a single command (npx wrangler deploy --temporary) without creating an account — the deployment provisions an ephemeral project for 60 minutes and outputs a claim URL to convert it to a permanent account.[13]Simon Willison Simon tested it on an HTTP redirect resolver tool and reported it "worked as advertised." He notes the AI agent framing is optional — this is useful for any developer who needs rapid, zero-friction deployment.

Read more

The 60-minute ephemeral window is designed for demos, prototyping, and quick sharing. If the deployer wants to keep the project, they follow the claim URL to attach it to an existing account or create a new one. The feature targets the friction point of "I want to share this prototype right now without going through account setup." Simon specifically pushes back on the AI agent marketing angle — pointing out this removes friction for all developers, not just automated agents.

Tools: Cloudflare Workers, Wrangler
AI Models
AICodeKing

Gemma 4 12B Agentic Fine-Tune: Promise with Caveats

A fine-tuned Gemma 4 12B model optimized for agentic coding tasks — released as GGUF under Apache 2.0 — went viral on Hugging Face with 6,000+ downloads in its first days, claiming 3.5x better tool-use performance than the base model on the Tau 2 Bench Telecom eval (15% → 55%).[14]AICodeKing In practice, the reviewer found it reliable for one-shot tasks but inconsistent in multi-step agentic workflows — prone to token repetition, strange output, and incomplete tool-use chains. The creator is already working on V3.

Read more

Available in Q3KM (5.7 GB), Q4KM (6.87 GB recommended), Q6K (9.11 GB), and Q8 (11.8 GB) GGUF variants — runnable via llama.cpp, Ollama, or LM Studio. Required sampling parameters: temperature 1.0, top-P 0.95, top-K 64, repetition penalty 1.1. The Tau 2 Bench eval is a small local eval (20 tasks), not directly comparable to the official leaderboard, so the 3.5x claim should be read cautiously.

"When it works, the good outputs can be quite good, but the consistency is not there yet."
Tools: Gemma 4 12B, Ollama, LM Studio, llama.cpp, Zed
AI Future
AI Search

GPT Discovers a Real Chemistry Improvement in the Lab

OpenAI integrated GPT with the Maria high-throughput automated chemistry platform and demonstrated a near-autonomous AI chemist that identified TEMPO as a novel additive significantly improving the yield of the Chan-Lam reaction — a method for forming carbon-nitrogen bonds important in drug synthesis — confirmed in real lab experiments.[4]AI Search GPT generated research hypotheses, designed experiments, analyzed results, and proposed follow-ups; human chemists provided steering and performed physical lab operations.

Read more

The Chan-Lam reaction forms C–N bonds used in medicinal chemistry, particularly for sulfonamide drug groups. GPT was given the open-ended goal of improving reaction yield, then systematically proposed and evaluated oxidant additives. TEMPO outperformed all other oxidants tested. A full technical paper was released. The workflow compressed what is typically a slow, expensive experimental cycle into a directed AI-human collaboration — with GPT doing the hypothesis and experiment design, Maria handling the physical automation, and the chemist team providing domain corrections.

Tools: GPT, Maria
Industry
AI Search

Sony Ace Beats a Human Pro; AGI Bot A3 Does It on Two Legs

Sony's autonomous table tennis robot "Ace" defeated a professional human player by detecting ball spin in real time, applying its own spin, and placing shots adaptively rather than just returning — mounted on a rail for rapid lateral movement.[4]AI Search In parallel, AGI Bot's A3 humanoid played table tennis while standing on two legs, powered by the Spike Ping Pong algorithm from Peking University — requiring simultaneous whole-body balance and ball reaction at millisecond precision.

Read more

Sony Ace detects ball spin — calculating rotation and axis of spin in milliseconds — and applies its own spin when striking, enabling it to actively force errors rather than passively defend. The rail mount enables the precision lateral movement needed for pro-speed rallies without the balance overhead of bipedal locomotion. The reviewer notes this is by far the best table tennis robot currently demonstrated.

AGI Bot A3 is a general-purpose humanoid designed for diverse tasks; the table tennis capability is a secondary demonstration. The Spike Ping Pong algorithm enables 10x faster vision response and millimeter-level precision for continuous rallies, trajectory tracking, and whole-body motion planning — all while maintaining balance on two legs.

AI Models
AI Search

Alibaba Logos: One AI for Proteins, Molecules, and Materials

Logos (Language of Generative Objects) is an open-source unified AI model from Alibaba's Tongyi Lab that handles proteins, small molecules, materials, antibodies, and chemical reactions within a single shared architecture — treating scientific domains the way a language model treats text.[4]AI Search Three model sizes (1B to 8B) are released under Apache 2.0; the 8B model is 16 GB and runs on most consumer hardware. It outperforms domain-specific competitors across ligand design, material generation, protein editing, and antibody design benchmarks.

Read more

Logos tokenizes molecular and biological data structures and applies the same architecture used for NLP, enabling a single model to predict and generate across protein design, antibody design, ligand design, material generation, and reaction prediction — tasks that previously required separate domain-specific tools. The Apache 2.0 license makes it commercially usable without restrictions, unlike some competing scientific AI models with research-only licenses.

Industry
AI Search

Midjourney Pivots to Full-Body Medical Scanning Spa

Midjourney announced "Midjourney Medical" — a full-body imaging system that immerses patients in a pool surrounded by ~500,000 sand-grain-sized ultrasonic sensor elements to reconstruct a 3D internal body map in 60 seconds — launching as a spa experience in San Francisco in 2027.[4]AI Search Rather than pursuing FDA clinical-scanner approval immediately, Midjourney is launching with hot tubs, saunas, cold plunges, and scanning pools — starting with body composition maps while seeking expanded approvals over time.

Read more

The sensors send ultrasound waves through the body from multiple angles simultaneously — similar to echolocation — and the returning wave patterns reconstruct a 3D internal map, processing terabytes of data per second. Target scan time is 60 seconds versus much longer MRI sessions, at a fraction of the cost. By 2028 they aim to scale to more cities and deploy a second-generation scanner with custom silicon. Regulation is the core constraint: FDA medical diagnostic approval requires demonstrating clinical validity, so the spa route sidesteps that while building evidence.

AI Tools
AI Search

Open-Source Video AI: World Models, Editors, and Style Transfer

A batch of open-source video and creative AI releases this week: DreamXWorld (5B interactive world model trained on Unreal Engine + real video, generates explorable persistent environments); Permavid (video editor with dual memory banks — appearance and 3D structure — solving AI video's consistency drift problem); LTX Trainer 2 (official fine-tuning package for LTX 2.3, the leading open-source video model with native audio); TeleStyle V2 (flexible style transfer built on Qwen Image Edit, handles realistic-to-realistic and arbitrary style combinations); BuGu Image (Apache 2.0 image generator and editor).[4]AI Search

Read more

DreamXWorld

Takes movement, camera control, or event instructions and outputs environments that stay consistent over hundreds of frames. The 5B model (~21 GB) runs on consumer hardware; a 14B version is planned. Trained on a mix of Unreal Engine gameplay and real-world video, outperforming World and HY World Play in consistency benchmarks.

Permavid

Separates appearance memory from 3D structure memory in two parallel banks — allowing global style changes to persist without destabilizing structure, and local edits to persist without affecting surrounding content. 29 GB model using CogVideoX 1.2.1 and Vace as base. Full dataset (400+ GB) also released.

LTX Trainer 2

Official LoRA training package for LTX 2.3. Supported workflows: video extension, inpainting, outpainting, text-to-audio, audio extension, video-to-video. Quick start guide and dataset preparation instructions included.

BuGu Image

Apache 2.0 — commercially usable without restriction. Base model (text-to-image) and edit model, each 20 GB full or 10 GB FP8. Turbo 4-step variant available. ComfyUI support already available. Key differentiator is licensing; photorealism quality trails Z Image and Flux Cline.

Tools: DreamXWorld, Permavid, LTX Trainer 2, LTX 2.3, TeleStyle V2, BuGu Image, Kling Omni Director, Qwen Image Edit, ComfyUI

Sources

  1. Newsletter The Rundown — Jun 17 + Jun 18, 2026
  2. YouTube The 5-Minute AI Weekly Recap: Realignment Week — The AI Daily Brief, Jun 21
  3. YouTube GLM 5.2 in 9 Minutes — Developers Digest, Jun 21
  4. YouTube New robot waifus, GLM 5.2 craze, AI spas, new world models: AI NEWS — AI Search, Jun 21
  5. Newsletter The Batch Issue 358 — DeepLearning.ai, Jun 19
  6. YouTube TypeScript is not TypeScript anymore... — Better Stack, Jun 21
  7. YouTube How the most AI-pilled product team builds products | Fiona Fung — Lenny's Podcast, Jun 21
  8. YouTube Every product of the future will be a living system — Ronak Malde, Trajectory.ai — Latent Space, Jun 21
  9. YouTube You Can't Run AI Agents Without This — AI News & Strategy Daily / Nate B Jones, Jun 21
  10. Blog Samsung Electronics brings ChatGPT Enterprise and Codex to all employees — OpenAI, Jun 21
  11. YouTube 6 Things to Know about AIE World's Fair 2026 — AI Engineer, Jun 21
  12. Blog sqlite-utils 4.0rc1 — Simon Willison, Jun 21
  13. Blog Temporary Cloudflare accounts — Simon Willison, Jun 21
  14. YouTube Gemma Coder (Fable Distilled): WHAT.. Fable LIKE OUTPUT from a 12B MODEL? — AICodeKing, Jun 21