Tech / AI / IT Monitor

Executive Summary

Anthropic's Fable 5 model dominated the cycle, returning after a brief nerf that reduced its benchmark performance, with developers reporting extraordinary capabilities for coding, web design, and even one-shot fine-tuning pipelines. NousResearch launched Hermes Agent v0.18.0 featuring a novel "Mixture of Agents" architecture, while the AI Engineer World's Fair in San Francisco served as the physical nexus for product demos, keynotes on agent memory and world models, and hardware showcases. On the hardware front, NVIDIA's DGX Spark is being actively benchmarked against Apple Silicon, and a sustained discourse emerged around US chip export controls inadvertently accelerating China's domestic AI stack—highlighted by Meituan's open-source LongCat-2.0 trained entirely on Chinese chips.

Key Events

• Anthropic releases/nerfs/re-releases Fable 5: Claude's most capable model was pulled, reportedly nerfed on benchmarks, then restored within 24 hours, with developers calling it a "reality compression algorithm" for thought-to-code translation. → link

• NousResearch ships Hermes Agent v0.18.0 "Judgement Release": Introduces Mixture of Agents as a first-class virtual model, /learn command, journey visualization, and support for Fable 5, Sonnet 5, and Gemini Vertex. → link

• AI Engineer World's Fair (Day 2): Keynotes from Chris Manning and Abhishek on sandboxing/world models; Logan Kilpatrick on generative UI with Cerebras; OpenAI booth featuring Codex reset distribution; Local AI Summit announced. → link

• DGX Spark benchmarks vs Apple Silicon: Ivan Fioravanti published Qwen 3.6-27B benchmarks showing DGX Spark wins on prefill/prompt processing, Apple Silicon wins on decoding; DGX Spark also tested with ComfyUI + Krea 2. → link

• DeepSeek DSpark speculative decoding lands in vLLM: First DSpark speculator for a non-DeepSeek frontier model (GLM 5.2) now runs natively in vLLM. → link

• "Freedom of Intelligence" hits WSJ: sqs advocates that government and Anthropic should not decide what level of intelligence is illegal for citizens, gaining mainstream press coverage. → link

• China AI self-reliance discourse: Meituan's LongCat-2.0 (1.6T parameters, trained on domestic Chinese chips, open-sourced) cited as evidence US export controls backfired by accelerating China's full-stack independence. → link

• Seedance 2.5 video model upgrade: Outputs noticeably more realistic than Seedance 2.0 on same prompts. → link

• Amp adds Fable mode: sqs ships claude-fable-5 as an Amp plugin mode, enabling 75%+ parallel agent development in orbs. → link

Analysis

The most striking pattern is the sheer velocity of Fable 5's adoption and cultural impact—within hours of its restoration, developers had one-shot fine-tuning pipelines, Mac apps, iOS features, and landing pages, suggesting Anthropic has achieved a step-function in thought-to-code translation. However, the nerf-then-restore sequence and user reports of quantized models emitting random Bengali mid-task point to a tension between capability and cost that labs are still navigating.

The hardware conversation is bifurcating: DGX Spark offers out-of-the-box CUDA compatibility with memory bandwidth limitations, while Apple Silicon requires MLX migration but excels at inference. Meanwhile, the export control discourse is shifting from speculation to evidence—China's open-source 1.6T parameter model on domestic chips is the first tangible proof that restrictions catalyzed self-reliance rather than suppression.

Agent tooling is rapidly consolidating: Mixture of Agents (Hermes), orchestrator-minion patterns (GLM + Fable), and speculative decoding (DSpark) all point toward architectures that route work across models of varying cost and capability. Watch for whether these multi-model patterns become standardized primitives or remain bespoke configurations.

Tweet Feed

Anthropic / Fable 5

@KingBootoshi · 2026-07-01T22:57

the best way i can describe fable is a 100x increase for navigating my world of ADHD [...] i consider it the greatest reality compression algorithm i have access to right now → link

@KingBootoshi · 2026-07-01T21:34

HOLY FUCK FABLE 5 ONE SHOT MY GEMMA 4 QB DUCK FINE TUNE [...] I AM GOING TO MAKE QB REAL → link

@KingBootoshi · 2026-07-01T23:29

asked Fable 5 to create a mac app for automating the full training pipeline, one shot [...] basically an AIO for training branded neural net mascots → link

@MilksandMatcha · 2026-07-02T17:11

sarah and dom on set pt. 2 — gpt 5.6, no more hyphens in OAI model naming?, all hail computer use → link

@jezell · 2026-07-02T15:08

RT @bridgemindai: FABLE 5 CAME BACK NERFED. We re-ran the July 1st version of Claude Fable 5 on BridgeBench. The results are brutal... → link

@sqs · 2026-07-01T20:36

Fable is available again in Amp: amp plugins add --auto-update @amp/fable-mode / amp --mode claude-fable-5 → link

@MengTo · 2026-07-02T07:26

I forgot how good fable 5 is at creating landing pages. It understands webgl, scroll behaviors and text animations so well with simple prompts. → link

@RayFernando1337 · 2026-07-02T16:20

I went to sleep and woke up with huge iOS features ready to go: onboarding, research for pricing, adding a paywall, ai proxy, and fixing multiple bugs. [...] unlocked with Fable 5 + Cursor Multitask. → link

@LinusEkenstam · 2026-07-01T20:37

fever fomo. need to rest. but damn fable. → link

@ivanfioravanti · 2026-07-01T19:53

Fable 5 is back! Let's use it till the next stop! → link

@ivanfioravanti · 2026-07-02T03:07

I think we can pass on Fable 5 for now 🤷🏻‍♂️ → link

@sudoingX · 2026-07-02T13:25

what quant of sonnet 5 are you serving over here @AnthropicAI? asked it for code, it stopped mid task to say "ভাই" at me and left a shell running. → link

@alexinexxx · 2026-07-01T20:23

Genesis 2:2 — And Anthropic said, "Let there be Fable." And there was Fable. And everyone thought that it was good. → link

@QuinnyPig (via @jezell) · 2026-07-02T04:58

Fable 5 is the most expensive Opus model router ever constructed. → link

Hermes Agent / NousResearch

@Teknium · 2026-07-01T20:24

Hermes Agent v0.18.0 is here. The Judgement Release. — Mixture of Agents as a first class virtual model, /learn to teach your agent anything, Journey to visualize how your agent learned, Gemini Vertex support, Fable 5/Sonnet 5/Fugu available → link

@Teknium · 2026-07-02T17:46

🤔🤔 → link

@Teknium · 2026-07-02T10:22

Interesting setup for our new Mixture of Agents feature! → link

@shannholmberg (via @Teknium) · 2026-07-02T17:45

what is the mixture-of-agents feature in Hermes Agent — normally you pick one model and trust its single answer, but mixture... → link

AI Engineer World's Fair

@swyx · 2026-07-02T06:07

for what it's worth, i only invite double-length track keynotes when I'm very sure both speaker and content deserve it. Today @chrmanning and @abshkbh did double duty on sandboxing and world models. → link

@swyx · 2026-07-02T18:50

very proud that the biggest applause line in the AIE keynotes this year was normalizing men talking about their feelings and mental health in hypergrowth → link

@MilksandMatcha · 2026-07-02T16:53

Generative UI, built for you on demand, principally requires fast inference. Logan Kilpatrick @OfficialLoganK on how AI could reshape software: agents handle transactional tasks, while new interfaces unlock richer experiences. Speed changes the interface. Presented with @cerebras → link

@NaderLikeLadder · 2026-07-01T21:44

Local AI Summit is tomorrow at AIE World Fair. Kicking off w/ a Local AI & OSS State of the Union panel at 10:45am. We'll demo GLM 5.2 running in the room on a DGX Station. → link

@alexocheema · 2026-07-02T16:40

Setting up with @NVIDIAAI — Local AI Summit, Room 2009 at @aiDotEngineer. Demos: Running GLM 5.2 on DGX Station, Running Nemotron 3 Ultra on 4 x DGX Spark → link

@steipete · 2026-07-01T21:31

Asked codex to download+transcribe all sessions from @aiDotEngineer and tailor them to my interests. → link

@TheAhmadOsman · 2026-07-02T01:43

THE LOCAL AI SUMMIT — AI ENGINEER WORLD FAIR IN SF — TOMORROW 10AM - 4PM — BE THERE → link

@MilksandMatcha · 2026-07-02T05:05

Token billionaires lounge — must apply and submit token spending receipts → link

DGX Spark / NVIDIA Hardware

@ivanfioravanti · 2026-07-02T09:53

Qwen3.6-27B MTP Context Benchmark on DGX Spark, M3 Ultra and M5 Max 🔥 — DGX Spark is the winner on Prefill/Prompt Processing. Apple Silicon on Decoding/Text Generation → link

@ivanfioravanti · 2026-07-01T20:32

DGC Spark Qwen3.6-27B-NVFP4 by @NVIDIAAI vLLM — Fast and furious! 🔥 [full benchmark numbers] → link

@ivanfioravanti · 2026-07-02T18:57

The thing I love most of DGX Spark is the fact that everything runs out of the box (thanks to Cuda), while on Apple Silicon I have to migrate models to MLX. But memory bandwidth is really too low on DGX. → link

@ivanfioravanti · 2026-07-02T14:40

DGX Spark + ComfyUI + Krea 2 Turbo ~14.5 secs per image, not bad and great quality! → link

@ivanfioravanti · 2026-07-02T15:27

Honestly I was thinking the same right now. At the end you have 128GB of RAM but everyone with a single DGX is running models can all stay within 24GB of a 3090. What am I missing? → link

@ivanfioravanti · 2026-07-01T19:12

Learning curve on DGX Spark is not trivial, because there are multiple ways to do things, in Apple Silicon there are just two: MLX (and derivatives) or llamacpp. In the Spark's world there is everything! 🤯 → link

@Ex0byt · 2026-07-01T19:31

the quality of innovation coming out of Nvidia AI has been dizzying of late.. Hope others start to incorporate the methods into their models. → link

China / Export Controls / Open Source AI

@sudoingX · 2026-07-02T09:12

the US banned its best AI chips from china to keep them years behind on AI. i think this is going to go down as one of the dumbest own goals in the history of tech. [...] meituan [...] just dropped an AI model that goes toe to toe with the best out of silicon valley. LongCat-2.0. 1.6 trillion parameters. trained on domestic chinese chips, zero nvidia anywhere. then they gave it away, open source. → link

@sudoingX · 2026-07-02T10:06

the export ban didn't stop china's ai. it started china's chip industry. dario wrote the whole essay arguing export controls would keep china behind. china answered with a 1.6T open model trained on its own chips. → link

@sudoingX · 2026-07-02T10:06

the day a chinese card gives me 48gb for what 12gb costs today, i'm not asking where it shipped from. cheap compute is the only ideology i have. → link

@sudoingX · 2026-07-02T11:35

dario is the man who hid obfuscated tracking code in your terminal for three months and only pulled it when he got caught. this is the man who wants to keep you safe. → link

@TheAhmadOsman · 2026-07-02T08:25

Closed frontier labs are more efficient and incentivized when Opensource labs exist. Doesn't that make Anthropic anti-capitalism? → link

@ClementDelangue (via @victormustar) · 2026-07-02T14:33

Lots of people are advocating for more American open-source models these days which is amazing but very few people do... → link

@maximelabonne (via @victormustar) · 2026-07-02T16:03

DeepSeek effect → link

Freedom of Intelligence

@sqs · 2026-07-02T01:32

Freedom of intelligence in the WSJ! Government and Anthropic should not decide what level of intelligence is illegal for citizens to access → link

@sqs · 2026-07-02T17:53

What would be needed for you to trust (and not babysit) an agent with access to the gcloud or aws CLI? Your work email? Your personal email? → link

@0xSero (via multiple) · 2026-07-02T17:14

Are you worried about your right to access intelligence? — local ai, open weight models, access to US frontier → link

DeepSeek / vLLM / Speculative Decoding

@vllm_project (via @ivanfioravanti) · 2026-07-02T05:01

🚀 @deepseek_ai's DSpark speculative decoding now runs natively in vLLM! — a semi-autoregressive drafter that... → link

@mgoin_ (via @victormustar) · 2026-07-02T16:49

GLM 5.2 DSpark preview is here! ✨ This is the first DSpark speculator for a non-DeepSeek frontier model... → link

@sudoingX · 2026-07-02T08:05

how do you load balance 256 experts in a 671B model? > deepseek's way: a bias term and a sign function. expert overloaded? -0.001. underloaded? +0.001. trillion dollar industry. thermostat logic.👌 → link

Codex / OpenAI / Developer Tools

@jxnlco · 2026-07-02T16:36

About to use codex computer use to control my iPhone via screen mirroring check find my to see who's around me and texts them. → link

@gdb · 2026-07-01T23:54

Codex for making a personalized daily digest: → link

@steipete · 2026-07-01T21:56

Pointed codex at some Twitter feedback on the OpenClaw iOS app and it did a first improvement pass. [...] It uses computer use to add before/after screenshots, as there's no GitHub API. → link

@sqs · 2026-07-02T18:08

"____ Frontier Corporation" is the new cool company naming convention. 1. Amp Frontier Corporation 2. Microsoft Frontier Company → link

@sqs · 2026-07-02T17:13

You, too, can take your development process from local to ~75%+ in orbs, infinitely parallel and not tied to your laptops. → link

@kunchenguid · 2026-07-02T18:08

just released lavish-axi v0.1.35 - it now supports exporting and publishing your html artifacts! "export" gets you a standalone html. "publish" hosts it... → link

@charliermarsh (via @jezell) · 2026-07-02T05:11

Codex made uncached, complex resolutions (like Transformers) in uv >40% faster through more efficient request scheduling... → link

@sudoingX · 2026-07-02T13:41

okay now claude code input cursor is broken too @bcherny wtf dude! first it's dropping bengali mid task, now i can't even type. claude code is having a day. → link

@thdxr · 2026-07-01T23:37

given how expensive (and slow) fable is we're trying to use it with the orchestrator + minion pattern (in this case GLM) — primary agent delegates all work and spawns them as background subagent → link

Agent Architecture / Multi-Model Patterns

@thdxr · 2026-07-01T20:42

the more experienced engineering put into a primitive, the more you can actually vibe code with it. i don't think we have all the primitives we'll ever need → link

@nicolaygerold (via @sqs) · 2026-07-02T16:36

Once we introduced compaction, our read_thread tool started to fall apart. One thread that broke it: 68 compactions, 21... → link

@sqs · 2026-07-02T00:14

I thought agents would be running more of my life by now → link

@hnasr · 2026-07-02T14:02

When a client process that is connected to a postgres instance is killed, what happens to the backend postgres process and all the locks it holds? → link

Hardware / Infrastructure

@cooltechtipz · 2026-07-02T15:21

AI chips depend on advanced packaging technologies like CoWoS, Foveros, and EMIB to connect processors with HBM. Demand has grown so quickly that packaging capacity is now one of the biggest bottlenecks. → link

@cooltechtipz · 2026-07-02T13:28

InfiniBand vs Ethernet → link

@FrameworkPuter · 2026-07-02T15:33

We're very happy to see Linux (and BSD) adoption continue to grow as a % of new users picking up Framework Laptops and Desktops, even as our sales volume continues to increase. → link

@jezell · 2026-07-02T18:57

Found a bug in containerd today. Living on the edge is fun. → link

Developer Tooling / Open Source

@kunchenguid · 2026-07-02T02:58

by popular ask i'm making my next YT video where i'll build all the configuration in my dotfiles from scratch and explain how everything works → link

@kunchenguid · 2026-07-01T23:56

i hope 2026 is the last year where we still have to manually click through any website to set things up — google cloud and apple app review are the two repeated offenders → link

@thdxr · 2026-07-01T20:52

update: everyone is installing opencode2 bots on their machines and they're invading the team chat → link

@steipete · 2026-07-02T04:06

Never thought I give @Steve_Yegge a shoutout. He was just early, like most visionaries. Now everyone is building factories. → link

@steipete · 2026-07-02T00:22

I'm looking for a semi-private hack space for a few days in SF for me and some of the OpenClaw maintainers. We need to cook. → link

@RayFernando1337 · 2026-07-02T00:27

iOS 27 will unlock a lot of great apps using local ai. → link

@KingBootoshi · 2026-07-02T03:14

qb's RLHF screen prototype → link

@jsuarez · 2026-07-01T19:33

Awesome work by our fintern! Multitask, multiagent racing + swarming on 1 GPU in 15 seconds. → link

Cerebras / Fast Inference

@cerebras (via @MilksandMatcha) · 2026-07-02T02:22

"If you knew you could get that many tokens, you would build different products." — Logan Kilpatrick on Cerebras speed enabling generative UI → link

@cerebras (via @MilksandMatcha) · 2026-07-01T23:43

We gave two agents the same task: "Find images matching this description." Both use Gemma 4 31B. One runs on Cerebras. → link

@victormustar (RT @googledevbr) · 2026-07-02T12:26

imagina a qualidade do Gemma4 31B em mais de 1000 tokens/s? [Gemma 4 31B running at >1000 tokens/s] → link

Local AI / On-Device

@slimcat0101 (via @victormustar) · 2026-07-02T16:44

Double down on this. 💯 Take OCR and document parsing for instance. Running a giant monster model is pure overkill. → link

@0xSero (via multiple) · 2026-07-02

Are you worried about your right to access intelligence? — local ai, open weight models, access to US frontier → link

@rudrank (via @RayFernando1337) · 2026-07-02T03:58

FoundationModelsKit 2.0.0 is out! I pulled the reusable bits back out of Foundation Lab after working with OS 27 → link

Seedance 2.5

@bearlyai (via @TrungTPhan) · 2026-07-02T18:03

Seedance 2.5 video outputs look much more realistic than Seedance 2.0 on same prompts → link

Miscellaneous / Commentary

@sqs · 2026-07-02T18:47

RT: Part 2, w @ankrgyl, @bernhardsson, @Sirupsen, @jerryjliu0, @travers00, @Steve_Yegge, @addyosmani → link

@gdb · 2026-07-01T22:14

you can just reset rate limits on things → link

@Levelsio · 2026-07-02T12:49

Why Claude keeps telling me to connect MCP to Google Drive etc? → link

@KingBootoshi · 2026-07-02T04:38

in self induced ai psychosis rn → link