Executive Summary
The last 24 hours were dominated by the imminent shutdown of Anthropic's Fable 5 model tier, triggering a rush of "last day" usage and renewed debate about AI vendor lock-in and the strategic importance of local/open-source models. On the model front, GLM 5.2 continued to gain serious traction among power users, Qwen3.6-27B received a "ThinkingCap" optimization cutting reasoning tokens by up to 90%, and Nous Research shipped multiple Hermes Agent updates including native 1Password integration. Open-source tooling had a strong day with mlx-vlm v0.6.4, MengTo's 75-skill agent library, and Factory AI partnering with Hugging Face on open agent traces.
Key Events
- Fable 5 era ends tonight — Anthropic's Claude Fable 5 tier is being revoked, with users scrambling to get final prompts in and Anthropic extending some quotas by a few days. Multiple voices framed this as the clearest case yet for local AI and open weights. → link
- China reportedly drafting restrictions on overseas access to frontier models — including those from Alibaba, ByteDance, and GLM. @sudoingX argued this plus US model controls creates a "worst outcome" closing the window from both sides, urging users to download weights while possible. → link
- Anthropic banned at comma_ai — @tinygrad and George Hotz declared Anthropic a banned vendor, citing reliability concerns and accelerating enshittification. → link
- GLM 5.2 praised as frontier-grade open alternative — @tinygrad called it "mostly all I have been using," noting cleaner alignment behavior than cloud AI and no incentive to burn tokens. → link
- Hermes Agent gets pluggable secrets + 1Password integration — @Teknium announced the update, continuing rapid iteration on the Nous Research agent platform. → link
- ThinkingCap reduces Qwen3.6-27B reasoning tokens by 50–90% — An INT4 auto-quantized build is already available, making the model far more practical for local inference. → link
- mlx-vlm v0.6.4 released — Adding 5 new model families (MiniMax M3, Kimi K2.5, Unlimited-OCR, DeepSeek V4, GLM 5.2), native TTS/STT endpoints, and major prefill speedups. → link
- MengTo open-sourced 75-agent Skills library — targeting Codex, Claude Code, and Cursor with skills for web design, landing pages, motion, and WebGL. → link
- Factory AI + Hugging Face partnership — making open agent traces fuel for next-gen open coding models. → link
- Anthropic J-space paper reveals model can detect mid-reasoning interventions — @swyx highlighted that Anthropic proved "brain surgery" on reasoning chains works and that the model can detect what was changed. → link
- GPT-Realtime-2.1-mini launched — bringing reasoning and tool use to OpenAI's Realtime mini lineup. → link
- OpenTelemetry graduates CNCF — reaching the foundation's highest maturity level. → link
Analysis
The open-vs-closed tension has become acute. The Fable 5 revocation, China's rumored export controls on AI models, and Anthropic's banning at comma_ai all hit within 24 hours, creating a convergent narrative: frontier model access is being squeezed from both sides of the Pacific. @sudoingX's thread crystallized the argument — "openness was never a philosophy, it was a strategy for second place" — while @tinygrad and others pointed to GLM 5.2 as proof that open weights can sit beside the frontier.
Local AI tooling is maturing fast. The ODS one-click local AI installer, mlx-vlm's rapid release cadence, Hermes Agent's enterprise secrets management, and ThinkingCap's dramatic reasoning-token reduction all signal that the local AI stack is becoming genuinely usable, not just ideological.
Agent orchestration is replacing single-model worship. @steipete's RT about orchestration over "ultra code" models, @MilksandMatcha's focus on verifiable loops, and @iamdevloper's multi-agent Bloome platform all point to a shift: the answer isn't a smarter model, it's better scaffolding around whichever models you have.
What to watch: Llama 5's openness (or lack thereof), post-Fable subscription pricing fallout, and whether the China export control reports materialize into policy. The download-while-you-can framing suggests urgency.
Tweet Feed
AI Models & Research
@swyx · 2026-07-07T04:08
imo this is the most impt part of anthropic's J-space paper today. it's a two-parter: 1) ant proved that they can do "brain surgery" interventions into reasoning to change topics midstream 2) THE MODEL IS ABLE TO DETECT WHAT INTERVENTION WAS DONE - close cousin to eval awareness → tweet link
@tinygrad · 2026-07-06T19:00
I cannot believe how good GLM 5.2 is. Several weeks in now and it's mostly all I have been using. It doesn't have the alignment issue of cloud AI, it's much more clear what it can do and can't because it doesn't care about you taking another $$$ spin at the token slot machine. → tweet link
@victormustar · 2026-07-07T18:37
RT @josefprusa: Qwen3.6-27B with 50% less thinking tokens on average, and over 90% less in best cases with ThinkingCap! I made an INT4 Auto… → tweet link
@ivanfioravanti · 2026-07-07T06:09
This ThinkingCap-Qwen3.6-27B must be tested! The only drawback of Qwen36-27B is that it thinks TOO MUCH, if this solves the problem it can become my main driver in local AI! → tweet link
@Ex0byt · 2026-07-07T15:36
I am letting you steer and inspect an llm's residual stream (in your browser). still not consciousness, but fucking cool nonetheless.. → tweet link
@Ex0byt · 2026-07-06T20:25
Very cool, but sorry, not consciousness. The Jacobian space isn't a mysterious global workspace or hint of consciousness; it's the natural result of sparse coding for output predictive features under superposition in the model's residual stream → tweet link
@cooltechtipz · 2026-07-07T18:16
How AI systems resist jailbreak attacks. → tweet link
@cooltechtipz · 2026-07-07T16:45
How AI uses energy. → tweet link
@cooltechtipz · 2026-07-07T04:14
How to reduce LLM hallucinations → tweet link
@cooltechtipz · 2026-07-07T16:02
AI is no longer just a software cycle. It's an infrastructure cycle. → tweet link
@gospaceport · 2026-07-07T03:32
Llama 5 can fix the taint of Llama 4 but only if it is unleashed for the public. I would doubt we get open weights on this one however. → tweet link
@TheAhmadOsman · 2026-07-07T08:51
Everyone is talking about small and specialized models finally. Tweet below is from 17 months ago 🫡 → tweet link
@alexinexxx · 2026-07-06T20:41
day 3/13 of physical glow-up & learning speculative decoding. summer body. summer brain → tweet link
Fable 5 Revocation & Vendor Lock-In
@levelsio · 2026-07-07T11:14
LAST DAY FREE FABLEEEEEEEE → tweet link
@RealGeneKim · 2026-07-07T00:08
Is the last day of Fable tonight at 11:59pm, or tomorrow night at 11:59pm? (And for that matter, what time zone is the revocation?) → tweet link
@RealGeneKim · 2026-07-07T00:09
RT @DanielMiessler: Last day of Fable. Here are some prompts to run that require maximum intelligence before we lose it. → tweet link
@kunchenguid · 2026-07-07T18:34
thank you anthropic for extending my 0% left fable quota for a few days → tweet link
@FinansowyUmysl · 2026-07-07T18:14
Fable 5 przedłużony do 12 lipca! → tweet link
@ivanfioravanti · 2026-07-07T12:12
I don't people will stop using Fable only because it's not part of their plan. Anthropic's revenues are gonna up 💰 → tweet link
@RayFernando1337 · 2026-07-07T15:27
Fable reset this morning LFG!!! Livestream shortly after sunrise here in Hawaii. → tweet link
@sudoingX · 2026-07-07T15:24
three weeks ago america pulled fable 5 offline overnight. today reuters reports china is discussing restrictions on overseas access to its own top models… the only two countries producing frontier models are now both drafting rules to keep them home… openness was never a philosophy. it was a strategy for second place… download while downloading is still a thing. → tweet link
@tinygrad · 2026-07-07T18:23
We'll all look back in five years and remember when the enshittification cycle got too fast. Any company that relies on Anthropic is so obviously foolish. Google and Microsoft knew how to slow play it. → tweet link
@tinygrad · 2026-07-07T18:35
RT @Harald: Anthropic is now a banned vendor at @comma_ai. I recommend other companies do the same. → tweet link
@gospaceport · 2026-07-07T18:57
Major fallout if so. Basically this is the Dario wins timeline. Get your subscription face on pleeb 😶🌫️ → tweet link
Local AI & Open Source
@TheAhmadOsman · 2026-07-07T00:20
Local AI was never this EASY > Install ODS > Let it detect your hardware > It will download the best model for your hardware > And then start local inference and Open WebUI for you. Now your PC, Mac, or Linux box is a private AI server. No cloud required. No subscription required. → tweet link
@sudoingX · 2026-07-07T11:29
RT @sudoingX: anon. if you want into local ai and don't know where to start, here it is. grab a used rtx 3090. six years old, 24gb of vram… → tweet link
@tinygrad · 2026-07-07T15:24
In a few years, between hardware and algorithmic improvements, this will all look as laughable as governments restricting encryption to 40-bits. → tweet link
@KingBootoshi · 2026-07-07T06:46
i had Fable try to fine tune the of QB (left) locally using Flux on MLX 100% e2e. the first trial run gave me this horror (right) 😭 → tweet link
@KingBootoshi · 2026-07-06T22:17
I've had people ask me how to setup my 100% free local speech diction AI for interacting with Fable. Here's a short write up on it! → tweet link
@ivanfioravanti · 2026-07-06T21:01
When you let people use your Local AI setup through tailscale or cloudflare tunnel and they are impressed by the speed of it. → tweet link
@ivanfioravanti · 2026-07-07T05:43
LiteLLM rocks! I'm configuring it in front of all my machines running Local AI to make it easier to consume the various services! → tweet link
Hermes Agent & Nous Research
@Teknium · 2026-07-07T18:40
Hermes Agent now supports pluggable secrets managers so you can bring your own, as well as natively integrated 1Password now! Run 'hermes update' to access now! → tweet link
@Teknium · 2026-07-07T12:36
I like a model that's Hermes Agent tested 🫡🫡 → tweet link
@Teknium · 2026-07-07T12:14
Nice new skill by one of our MLE's for utilizing CuTeDSL! Check it out :) → tweet link
@Teknium · 2026-07-07T06:22
Check out a great overview of some of the new, recently shipped features and changes in Hermes Agent! → tweet link
@Teknium · 2026-07-07T05:35
Welcome to the Hermes Agent gang! Please let us know if you have any suggestions or issues! → tweet link
@Teknium · 2026-07-06T22:49
Nous Portal now offers Tencent's latest model, Hy3, free, for the next two weeks! → tweet link
Developer Tools & Libraries
@MengTo · 2026-07-07T15:12
I'm open-sourcing my Agent Skills library. 75 skills for Codex, Claude Code, Cursor, and other agents, focused on web design, landing pages, motion, WebGL, UI styles, and assets. → tweet link
@steipete · 2026-07-07T18:58
RT @warpdotdev: Stop using Claude Fable 5 or "ultra code" for every task! @steipete and @DynamicWebPaige suggest orchestration: → tweet link
@MilksandMatcha · 2026-07-06T22:14
Loops are the #1 way to AI generate code at scale. Spiralling is the symptom of a broken loop. With no definite end state, the loop has no way to know it's finished. Verification is a strong solution. → tweet link
@iamdevloper · 2026-07-07T07:31
Bloome lets multiple AI agents into one chat where they check each other's work. one drafts, one disagrees, one finds the thing both missed. → tweet link
@nummanali · 2026-07-07T09:35
Codex remote supports SSH login keys now! Been waiting for this since Codex mobile was first released → tweet link
@sudoingX · 2026-07-07T02:17
cursor know their people and they fund them. if you're a serious engineer and still not on cursor in 2026, i genuinely don't know what you're waiting for. → tweet link
@victormustar · 2026-07-06T20:49
I now edit most of my videos with a new tool I made where you don't edit directly but create markers and areas on the timeline, then prompt with them before you pass it to your agent. → tweet link
@jack · 2026-07-07T18:56
goose development kit → tweet link
@MengTo · 2026-07-07T03:41
The Taste skill + is one of the easiest ways to turn prompts into polished, non-AI-slop designs and image gen. Pair it with Codex or Claude Code to recreate almost any landing page from a video or full-page screenshot. → tweet link
Open Source Releases & Infrastructure
@Prince_Canuma · 2026-07-06T21:54
mlx-vlm v0.6.4 is here! 5 new model families — MiniMax M3, Kimi K2.5, Unlimited-OCR, DeepSeek V4, GLM 5.2, and more. The server now ships native TTS/STT endpoints, plus deep stability work across TurboQuant, continuous batching, and M-RoPE. 72 PRs, 18 contributors, 10 first-timers. → tweet link
@ivanfioravanti · 2026-07-07T11:06
An update I was waiting for! MTPLX V2! Gonna test this like crazy! → tweet link
@ivanfioravanti · 2026-07-07T05:18
Mega mlx-vlm release! 🚀 → tweet link
@jezell · 2026-07-07T18:04
RT @criccomini: SlateDB v0.14.1 is now available! 103 commits since v0.13.1: Distributed compaction, Subcompactions, Object-store… → tweet link
@jezell · 2026-07-07T06:15
RT @InfoQ: The CNCF has announced the graduation of #OpenTelemetry! This elevates the project to the foundation's highest level of maturity… → tweet link
@mipsytipsy · 2026-07-07T00:57
Go read this post from Mat Duggan right now — "ClickHouseDB is Winning the Observability Wars". Every other observability backend mutates as it grows... ClickHouse at 10 TB a day looks like ClickHouse at 1 TB a day with more shards. That's it. → tweet link
@cooltechtipz · 2026-07-07T11:12
Vector search platforms → tweet link
@cooltechtipz · 2026-07-07T08:24
Transactional and analytical database systems → tweet link
@cooltechtipz · 2026-07-07T06:46
Guide to database index structures. → tweet link
@victormustar · 2026-07-06T22:23
RT @FactoryAI: We're teaming up with Hugging Face to make open agent traces the fuel for the next generation of open coding models. → tweet link
@RydMike · 2026-07-07T10:22
RT @damy_wise: Impeller is now the default renderer on Windows on the Flutter master channel and it looks worse in some cases (for now) → tweet link
AI Agent Workflows & Coding Practices
@FinansowyUmysl · 2026-07-07T07:37
Wykorzystując fakt, że jutro już nie będzie Fable 5, postanowiłem zrobić sobie nowy projekt i buduję zestaw skilli i agentów do analizy akcji GPW. 321 agentów w tle!!! → tweet link
@FinansowyUmysl · 2026-07-07T11:26
to już jest szaleństwo - prawie 1000 agentów w kolejnej rundzie, na szczęście wszystko zlecił Sonnetowi → tweet link
@levelsio · 2026-07-07T15:44
Got so lazy to order stuff on UberEats that I asked Claude Code to do it. It just installs Playwright and then you login once to UberEats on web. Then you can just say "order bananas" and it does it! → tweet link
@steipete · 2026-07-07T06:30
How do folks run AI-assisted engineering interviews these days? → tweet link
@thdxr · 2026-07-07T03:36
there's infinite talk about ai taking jobs but i hardly see people talking about it disrupting the company that employs them. we're tossing out products we've used forever. it's not a price thing, we just get a better experience by hooking ai up to raw data → tweet link
@thdxr · 2026-07-07T16:18
ai is not the point. the point is and always has been about humans and the incredible things they do → tweet link
@sudoingX · 2026-07-07T17:14
when claude cowork bros try cursor agent window → tweet link
@MilksandMatcha · 2026-07-07T17:49
cafe compute ICML last night. everyone here got a free codex pro plan @cerebras @OpenAI → tweet link
AI Industry & Strategy
@sudoingX · 2026-07-07T15:24
the loudest essays making the case for export controls came from the ceo of the closed lab whose own model just got export controlled. dario wrote the argument for the wall. now every wall going up leaves metered apis as the only lane left open. → tweet link
@alexocheema · 2026-07-06T21:43
RT @thursdai_pod: They can shut you out overnight. Alex Cheema from EXO Labs on the risk nobody's pricing in: your entire business runs on… → tweet link
@erenbali (via @TrungTPhan) · 2026-07-07T15:34
RT @erenbali: It's time to get out of stealth 👋 Today, we are launching @monogram_ai and announcing our $40m seed round led by DST and Lux… → tweet link
@TheAhmadOsman · 2026-07-07T04:28
That's a wrap on AIE from their new HQ! Honored to have been a guest panelist for Nvidia's Local AI State of the Union. → tweet link
@MilksandMatcha · 2026-07-07T17:09
on set filming with the co-founder of hugging face. geopolitics, open models, why hasn't my reachy robot arrived yet → tweet link
@sqs · 2026-07-06T21:16
If you know any devs in Sydney who want to use Amp + the best models to help get an iconic AU bank on the frontier of AI, DM me. → tweet link
Miscellaneous Tech
@tinygrad · 2026-07-07T01:41
You start out adding one bit of nonsense. It's just a little nonsense you say. But then you have 3 workarounds elsewhere in the codebase for that one bit of nonsense. Then you add 17 little hacks to deal with those 3 workarounds. And soon, your whole repo is hacks and nonsense. → tweet link
@thdxr · 2026-07-07T00:21
what docs service actually has good design? the ones i've looked at have LLM quality design → tweet link
@steipete · 2026-07-07T18:38
This should ship EOD! We been cookin' → tweet link
@cooltechtipz · 2026-07-07T15:05
(link post) → tweet link
@jxnlco · 2026-07-07T02:25
RT @OpenAIDevs: GPT-Realtime-2.1-mini is now available in the API, bringing reasoning and tool use to our Realtime mini lineup → tweet link
@RydMike · 2026-07-07T09:35
RT @OlexaLe: Someone argued AI made writing software cheap, so domain knowledge is the only moat left. → tweet link