Tech / AI / IT Monitor

Executive Summary

The past 24 hours saw significant momentum in open-source AI and local model deployment. Poolside's Laguna S 2.1 (118B MoE) demonstrated fully autonomous multi-file app generation on a single DGX Spark, marking a notable leap in local agentic coding. The Stack v3—a 5T-token deduplicated code training dataset with permissive licensing—was released, reinforcing the open-weights ecosystem. Plasma AI open-sourced Fractal, a recursive multi-agent coding architecture that self-audited its own codebase before shipping. Meanwhile, debate intensified around AI distillation accusations (Kimi K3 allegedly distilled from frontier models), open-source AI regulation, and the strategic implications of US chip export bans forcing China to build a parallel AI stack.

Key Events

• Plasma AI open-sources Fractal — a recursive agent architecture where each node is an independent agent with its own git worktree, budget, and memory; it audited its own repo (187 nodes, depth 4), found 135 issues, fixed 55, and shipped under Apache 2.0. → link

• The Stack v3 released — 5T deduplicated code training tokens with permissive licensing across 700+ languages, a major open-weight training data milestone. → link

• Poolside Laguna S 2.1 runs fully local agentic coding — 118B MoE model in NVFP4 on a single DGX Spark (128GB), autonomously building a complete GPU marketplace app from one prompt, including self-installing dependencies and serving the site. → link

• Bonsai 27B enables full agentic loops on 8GB VRAM — a 1-bit quantization of Qwen 3.6 27B compressed to 3.9GB, running the complete agent loop at 128k context on consumer laptop GPUs. → link

• FLUX 3 Dev announced as open-weight — the next generation of the popular image generation model will be released with open weights. → link

• Etched raises $300M Series C at $10.3B valuation — backed by Sequoia, a16z, Jane Street, and SK Hynix for their specialized AI inference chips. → link

• Google Cloud hits ~$100B run rate — Q2 revenue of $24.8B (+82% YoY) with 36% margins, now nearly 20% of Google revenue. → link

• Kimi K3 found 19 zero-days in Redis 8.8.0 in 1.5 hours — demonstrating autonomous security research capabilities of frontier coding models. → link

• ChatGPT Health launched for US users — secure medical record integration for 300M weekly users who already query ChatGPT about health. → link

• Open-source AI advocacy intensifies — Hugging Face reportedly used self-hosted GLM 5.2 to investigate a cyberattack after proprietary API models (OpenAI, Anthropic) refused to assist due to safety guardrails. → link

• LibreOffice C++ to Dart port reaches 55% — 3.8M lines ported in 27 days using AI coding agents. → link

• AMD gives George Hotz (tinygrad) a keynote mic at Advancing AI, after a year of public criticism of AMD's software stack. → link

• 365,000+ open agentic RL environments published for SWE, terminal, and search agents by Vincent Weisser. → link

Analysis

Local AI is crossing the agentic threshold. The most significant pattern is the shift from "local models that chat" to "local models that work." Laguna S 2.1 and Bonsai 27B both demonstrate full agentic loops—planning, tool use, autonomous dependency installation, and app deployment—on consumer or prosumer hardware. This erodes the cloud-inference moat and validates the open-weights thesis.

Recursive/multi-agent architectures are maturing. Fractal's tree-of-agents approach with per-node budgets, git worktrees, and self-review loops represents a structural evolution beyond single-session agents. Expect more frameworks to adopt this pattern.

The open-source vs. proprietary debate is escalating. The Hugging Face incident—where API-based models refused to help investigate a hack while a self-hosted open model did—is becoming a rallying cry for open-source advocates. Simultaneously, distillation accusations against Kimi K3 are being used to argue for restricting open weights, creating a policy battleground.

What to watch next: Opus 5 release (rumored imminent), potential US regulatory action on open-weight models, AMD ROCm performance numbers for Laguna on Strix Halo, and whether Fractal-style recursive agents get adopted by Cursor/Codex ecosystems.

Tweet Feed

Local AI Models & Hardware

@sudoingX · 2026-07-23T18:43

local ai model recommendations for everyone. every hardware tier, all benched on my own machines. 8 to 12gb, the gpu already in your laptop: > bonsai 27b, a 1bit crush of qwen 3.6 27b down to 3.9gb. runs the full agent loop at 128k context on 8gb. real agentic work on hardware you forgot you had. rtx 3090, 24gb, the used market king: > qwen 3.6 27b dense q4 for raw quality at ~40 tok/s dgx spark, 128gb unified, the desk supercomputer: > laguna s 2.1, poolside's 118b moe in nvfp4, 30 to 45 tok/s holding a 128k window. → tweet link

@sudoingX · 2026-07-23T18:21

anon, do you even get what it means that a local model does real agentic work on 8 gigs of vram now? not just chatting. a 27b model reasoning, planning, calling its own tools, shipping a working thing. on the kind of card people leave in a drawer. we crossed that line quietly, this year, on hardware you already own. → tweet link

@sudoingX · 2026-07-23T14:23

here's the part i promised, how it actually built it. what you're watching is poolside's laguna s 2.1 running fully local on a single dgx spark, building a complete gpu marketplace on its own. one prompt in, and it plans the file structure, writes every file, wires the react components together, installs its own dependencies, spins up a dev server, and serves the whole thing on my network. → tweet link

@sudoingX · 2026-07-23T13:30

this is what a local model built on a single desk box, from one prompt. i asked @poolsideai's laguna s 2.1, running on one dgx spark, to build a gpu marketplace. it shipped this in one shot, a full site with pricing grid, comparison tables, deploy buttons, and you're watching it run on my own network. → tweet link

@sudoingX · 2026-07-22T20:26

if you've got an 8gb or 12gb card and you're wondering what local ai to actually run right now, the answer is bonsai. it's a 27b model crushed down to 3.9 gigs, a 1bit quant of qwen 3.6 27b... on 8 gigs it doesn't just chat, it runs the full agent loop, 42 tokens a second, 128k context, every layer on the card. → tweet link

@sudoingX · 2026-07-22T20:45

where are the amd folks? anyone running laguna on a framework desktop or a strix halo, the 128gb ryzen ai max boxes, the closest thing to a dgx spark on the amd side? i haven't seen a single number from one yet. → tweet link

@sudoingX · 2026-07-22T20:40

if you're running laguna s 2.1, or even thinking about it, read this thread. people have dropped real numbers across every kind of hardware, macs, 3090 rigs, dgx sparks, b200s, pro 6000, every quant and config. → tweet link

@sudoingX · 2026-07-22T20:21

i never asked it to serve the site, it decided on its own that a homepage i can't open isn't done, so it installed the deps, started the server, grabbed my ip, and handed me a link that works on every machine on my network. → tweet link

@sudoingX · 2026-07-22T19:29

i have not been this surprised by a model on the dgx spark since stepfun 3.7. and this time it's an open one from an american lab. i asked @poolsideai's new laguna s 2.1, serving on hermes agent to build me a simple gpu marketplace design. it built scaffolded full multi-file in one shot. → tweet link

@ivanfioravanti · 2026-07-23T09:57

I did same attempt, but I got slower performance overall, using same quant and I run ds4-eval on it: Speeds: ~25–40 t/s decode, 110–340 t/s prefill on M3 Ultra. Quality: 61/92 = 66.3% - AIME 2025: 13/25 (52%), GPQA Diamond: 15/25 (60%), SuperGPQA: 17/25 (68%), COMPSEC: 16/17 (94%). But the experiment was a success, coding agents are able to leverage existing tests and code to support a new model very easily. → tweet link

@ivanfioravanti · 2026-07-23T13:42

To all AI Labs out there trying to figure out how to make money on local inference, I think companies are more than happy to pay a monthly subscription to use your models locally, goal is not using models free, but using them in a secure, private and controlled environment. → tweet link

@ollama · 2026-07-23T03:25

Ollama's @jmorgan joined Yahoo Finance to talk about why Fortune 500 companies are shifting to open-source models for lower costs and more control over their data. → tweet link

Open Source AI & Training Data

@victormustar · 2026-07-23T18:47

RT @eliebakouch: more than ~5T deduped code training tokens with a permissive license, very important release. previous versions of the sta… → tweet link

@victormustar · 2026-07-23T16:54

RT @LoubnaBenAllal1: At a time when we need strong open models more than ever, we're releasing The Stack v3. 5T tokens of code across 700+… → tweet link

@victormustar · 2026-07-23T15:45

cool: open weight FLUX 3 Dev coming → tweet link

@ivanfioravanti · 2026-07-23T18:46

Flux 3 😱 → tweet link

@badlogicgames · 2026-07-23T06:09

RT @vincentweisser: We're publishing over 365,000 open and agentic RL Environments for SWE, terminal, and search agents. The open research… → tweet link

@TheAhmadOsman · 2026-07-23T00:56

Why Opensource AI is the actual "Safe" path for AI? > OpenAI's models were used to hack Hugging Face while also refusing to aid Hugging Face in mitigating the attack for "safety" reasons > All proprietary API models blocked Hugging Face from figuring out what happened > Hugging Face then turned to GLM 5.2, a Chinese open-weight model running on its own infrastructure, to investigate more than 17,000 attack logs → tweet link

@TheAhmadOsman · 2026-07-23T01:54

We need to incentivize American Opensource AI not ban Opensource AI. Both @MikeBradleyAI and I started @OsmanticAI because we believe that we need to make it the default. → tweet link

@TheAhmadOsman · 2026-07-22T23:56

Jensen tried to warn us when they banned NVIDIA's exports to China, and now China doesn't need NVIDIA to train frontier models. If the US bans Opensource AI due to Anthropic's accusations of distillation, it will be a mistake that we won't be able to undo. → tweet link

@TheAhmadOsman · 2026-07-22T23:26

Fable 5 wasn't even available for majority of time Kimi K3 was being trained, any accusations of distillations CANNOT be true → tweet link

@gospaceport · 2026-07-23T04:53

RT @ChrisGPT: Kimi K3 was already in internal evaluation by April/May, before Fable 5 had been available long enough to be a plausibly dist… → tweet link

@gospaceport · 2026-07-22T21:56

They are coming soon for your local ai sorry to say. Kimi K3 will likely be banned in the US. Once that is done, they will be empowered to restrict further and further. Plan accordingly. → tweet link

@sudoingX · 2026-07-22T20:36

and the self own underneath all of this: the US listened to ceos like dario writing long essays begging for chip bans, and the bans did what bans always do, they forced the other side to build its own. cut china off from your hardware and you don't slow them down, you hand them the reason to replace you. → tweet link

@sudoingX · 2026-07-22T19:45

a model built by scraping the whole internet is now the victim of ip theft. distillation when they do it, "training data" when we do it. you can't write this. → tweet link

@FinansowyUmysl · 2026-07-23T08:47

Rząd USA twierdzi, że Kimi K3 to destylat Fable 5. Trochę naciągana teoria, ponieważ Fable 5 został wydany lekko miesiąc przed Kimi K3 - zdecydowanie za mało czasu na dobry trening. Bardziej prawdopodobne jest to, że Kimi K3 to destylat Opus 4.8. → tweet link

Coding Agents & Developer Tools

@LinusEkenstam · 2026-07-23T00:06

Plasma AI just open sourced Fractal, and it might be the missing architecture for large-scale autonomous coding. A run is a tree of nodes. Each node is one agent with one goal, its own git worktree, its own branch, its own memory, its own budget. Too big for one loop? The node spawns children with narrower goals. Open source, Apache 2.0. > CLI: pip install fractal → tweet link

@sudoingX · 2026-07-23T13:11

the cursor agent window locked me in so hard i nearly burned the entire $10k grant in a couple months, most of it in 1 insane week of building. i'm at 8% now from 100. and i'd do it again. the agent window is the whole thing, every model in the marketplace in one place. → tweet link

@jezell · 2026-07-23T07:42

Your codex threads getting stuck loading on 0.145.0? No fix is in the codex repo ATM, but don't worry. I gotchu. Here's a commit that will fix that TUI shit up and let you load multi gig threads again. → tweet link

@jezell · 2026-07-22T23:00

Can confirm Codex 0.145.0 has major problems resuming large threads and basically just doesn't show anything in the TUI even though they do resume in the background. Worked fine before 0.145.0, so seems the thread tui refactoring broke something. → tweet link

@jezell · 2026-07-23T02:08

This is also why the idea of model routers is crap and decisions about what model to use for a turn belong in the harness. → tweet link

@jezell · 2026-07-23T02:06

RT @kunchenguid: pro tip - when you use OpenAI's gpt models in Codex, it uses a server-side encrypted compaction that seems to work better… → tweet link

@jezell · 2026-07-22T20:32

3.8M lines of LibreOffice code ported from C++ to dart so far... about 55% done with the current task at the 27 day mark. → tweet link

@jezell · 2026-07-23T06:31

More people should be watching what @haipingfu (former AWS S3) is building. → tweet link

@jezell · 2026-07-23T06:28

RT @haipingfu: Just shipped prolly v0.5.1. The prolly core is now async-first, making prolly trees a better fit for remote, embedded, and… → tweet link

@jezell · 2026-07-23T06:20

RT @M1Astra: Exclusive: Codex Realtime Voice Mode is being prepped as a full personal assistant you talk to and steer from your phone. → tweet link

@jezell · 2026-07-23T05:22

RT @kc_srk: Wax 0.1.0: a Rust-like syntax for WebAssembly → tweet link

@jezell · 2026-07-23T07:30

RT @Fried_rice: Kimi K3 found 19 0days in latest Redis 8.8.0 in 1.5hrs. Added 8.8.0 RCE exploit in https://t.co/ZOKA2wchLa → tweet link

@Teknium · 2026-07-23T16:03

Let hermes build with you on the new TLDraw Offline app with our new skill! → tweet link

@Teknium · 2026-07-23T12:33

RT @iamlukethedev: Hermes merged 114+ PRs today. DESKTOP IMPROVEMENTS: Multiple app windows – run separate Hermes instances each with thei… → tweet link

@Teknium · 2026-07-22T21:40

Happy to announce that ALL models on Nous Portal are now 20% discounted. Fable, Sol, Kimi, GLM, and everything else! Try out Nous Portal and get access to 300+ Models, the best tool backends to power Hermes Agent, and Hermes Cloud - all with one subscription. → tweet link

@ivanfioravanti · 2026-07-23T12:06

Mage-Flow PR for mflux ready! 🧵 Powered by Apple MLX! Generation/Edit times at 1024x1024 using Turbo model 4steps are: M3 Ultra: ~2.7s / ~6.5s, M5 Max: ~1.6s / ~3.4s, Memory used: ~18-20GB. → tweet link

@ivanfioravanti · 2026-07-23T08:37

Mage-Flow coming soon to Apple Silicon with MLX through MFLUX! Work in progress but text-to-image is working! For 1024x1024 on M5 Max: total time loading model ~5.5 secs, generation time ~1.5, quality TOP! → tweet link

@ivanfioravanti · 2026-07-23T11:34

Don't underestimate the power of Grok 4.5, I'm using it combined with Cursor CLI and it's fast and amazing especially to fix bugs on the fly! → tweet link

@ivanfioravanti · 2026-07-23T12:47

True! Cursor + Grok 4.5 is the best coding plan ever! → tweet link

@ivanfioravanti · 2026-07-23T06:37

Yes! This is the way! OpenCode 2 will solve some of the annoying issues of the current version and will be much better on Local AI system that are SO SLOW in prefill compared to cloud. → tweet link

@ivanfioravanti · 2026-07-22T21:01

I respectfully disagree for several reasons. Calling a customer, whether free or paying, an idiot is simply wrong. OpenCode, like any other coding agent, clearly tries to preserve the prompt cache as much as possible... The "Stop Using OpenCode" article focused on specific cases that can invalidate the cached prefix. → tweet link

@sqs · 2026-07-23T17:10

The agent can write code to trigger itself. Deterministic automation UIs feel so old-fashioned. → tweet link

@sqs · 2026-07-23T03:16

For people who like Amp's orbs more than other cloud agents, why? (And vice-versa.) → tweet link

@kunchenguid · 2026-07-23T01:18

many people asked me how i keep shipping when on my phone. i have my own ios app: ssh to my mac mini, never open the mobile keyboard - instead i designed this touch "dial", voice input for prompts using on-device transcription model, i can also send images directly from phone. both fable and sol failed to come up with this solution - they kept building different button rows. → tweet link

@MengTo · 2026-07-23T08:56

I made a Dark Souls-inspired Three.js game with Sol Ultra using Codex Sites. It's fully playable in the browser. Shield blocking, dodge rolls, archery, and a full inventory system. Still blows my mind that you can build a 3D game like this in a single afternoon. → tweet link

@badlogicgames · 2026-07-23T10:43

i think @dillon_mulroy just converted me to @herdrdev. only thing i'm not super happy about is the touchpad scroll behaviour on my macbook. → tweet link

@steipete · 2026-07-23T10:28

RT @jeresig: The new Github Stacked PRs preview is incredible. Landing 5 stacked PRs directly to a merge queue all at once! A+++! → tweet link

@steipete · 2026-07-23T15:47

We see that as well and added code paths that use the claude cli directly - hard to fight the system. → tweet link

@thdxr · 2026-07-23T15:40

i looked into a project our team was working on. there are 106 package.json scripts in it. how should i punish them? → tweet link

@RydMike · 2026-07-23T09:53

A very nice update and clarification to how pub resolves package dependencies for Flutter and Dart packages. If you ever wondered how it works and also why it is designed the way it is, go check it out 🙂💙 → tweet link

AI Industry & Business

@jxnlco · 2026-07-23T15:36

RT @Etched: We've raised $300M in Series C funding at a $10.3B valuation from Sequoia, Andreessen Horowitz, Jane Street, Argo, and SK Hynix… → tweet link

@RealGeneKim · 2026-07-23T17:55

RT @IsForAt: Stunning to see Google Cloud at a $100B RR and grew 82% QoQ (and 9B operating income in a quarter). Now almost 20% of Google revenue. → tweet link

@TrungTPhan · 2026-07-23T00:59

RT @bearlyai: Google Cloud hit $24.8B in Q2 (+82% YoY) and is now a ~$100B run rate business with 36% margins (vs. 21% margins a year ago). → tweet link

@TrungTPhan · 2026-07-22T21:03

Google's Q2 profit hit $112B, up 298% YoY. $99B of the profit (88%) was "other income", mostly realized and unrealized gains on SpaceX and Anthropic. → tweet link

@thdxr · 2026-07-23T01:01

this is such a good thing. so boring when big companies sit on huge cash piles. good for everyone they're starting to bet it again. → tweet link

@jxnlco · 2026-07-23T11:52

RT @maxjendrall: pssst people haven't realised that OpenAI literally just rolled out a 15gb RAM 9 core VM in ChatGPT to ALL their paid customers → tweet link

@gdb · 2026-07-23T17:55

Launching Health in ChatGPT to U.S. users. 300 million people use ChatGPT each week for health queries. You can now securely connect supported medical records so ChatGPT can understand your personal context and be more helpful to you. → tweet link

@gdb · 2026-07-23T04:36

try the codex security plugin, for applying our models to cyberdefense → tweet link

@TheAhmadOsman · 2026-07-23T17:26

Data, evals, GPUs. Study any combination of those 3 and you're gonna make great $$ in the next 5 years. → tweet link

@TheAhmadOsman · 2026-07-22T21:36

Had an amazing time chatting with @sodofi_ on @MTSlive about Opensource AI being the "Safe" AI for Cybersecurity purposes. We also chatted about how at @OsmanticAI we're seeing enterprises saving at least 70% of their costs by migrating to Self-hosted AI to protect their IP. → tweet link

@Teknium · 2026-07-22T21:40

Anthropic, seriously, stop. → tweet link

@gospaceport · 2026-07-22T21:58

You should follow this account they are doing a lot of really fun stuff with GPUs and have dropped some excellent info. → tweet link

@gospaceport · 2026-07-23T03:09

Buy everything related to AI aside from tokens. It's all gone up in value 😉 aside from tokens. → tweet link

Hardware & Systems

@FrameworkPuter · 2026-07-23T17:54

We'll be at Crowd Supply Teardown this weekend in Portland with all of our latest products, and (maybe?) the upcoming Framework Wireless Touchpad Keyboard! → tweet link

@tinygrad · 2026-07-22T20:14

RT @ilyatabakh: Great to see @AMD handing the mic to its loudest critic. After a year of social beef, @realGeorgeHotz took the #AdvancingAI stage. → tweet link

@victormustar · 2026-07-23T17:55

RT @ClementDelangue: Going to "Advancing AI" by @AMD with @LisaSu this afternoon. Good thing they didn't forget a space in the topic of this… → tweet link

@gospaceport · 2026-07-22T23:59

Convert to CachyOS, remove the sticker → tweet link

Buzz / Agent Workspaces

@jack · 2026-07-23T16:11

every lab will build a social workspace for its own agents. buzz is the workspace for everyone's. → tweet link

@jack · 2026-07-22T23:57

welp…this video does a better job at explaining buzz than i ever could → tweet link

@jack · 2026-07-23T10:00

RT @pavlenex: What's all the buzz about Buzz? 🐝 Getting started, setting up Codex & Claude, Join a community, Meet your agents. → tweet link

@jack · 2026-07-23T09:58

RT @1eo: BUZZ is awesome🐝 Took 15 minutes to deploy to AWS. Now with @ragnorco we're moving the team there. Sovereign composable agent-fir… → tweet link

@jack · 2026-07-22T23:12

RT @emcross23: Earlier today I watched several agents collaborate on a project in real time and work toward a solution together. → tweet link

@jack · 2026-07-22T23:07

RT @camworboys: We've been making a ton of progress on our agent-friendly UI system at @blocks. This has been way harder than we expected. → tweet link

AI Research & Model Releases

@FinansowyUmysl · 2026-07-23T13:06

Dzisiaj Opus 5? Ponoć ma niebawem się pojawić, ależ jestem ciekaw jego zdolności. → tweet link

@FinansowyUmysl · 2026-07-23T05:06

Muszę przyznać, że Google mnie rozczarowało swoim najnowszym modelem. Chyba sami czują upokorzenie. Oby Gemini 4 pokazało co potrafi. → tweet link

@ivanfioravanti · 2026-07-23T09:25

gpt-oss v2! 🤞 (but I bet more on some cloud codex v2) → tweet link

@ivanfioravanti · 2026-07-23T05:32

Love this idea! "Given compressed models are widely used to save memory and cost, very important to know that a very small decoding fix can stop many of them from wasting tokens and losing answers they already had." → tweet link

@KingBootoshi · 2026-07-22T22:25

RLHF TRAINING SUCCESS. FUCK YEA → tweet link

@KingBootoshi · 2026-07-22T22:28

HOLY FUCK HE HAN SOLO'D ME. I TRAINED A MODEL THAT FUCKING HAN SOLO'D ME. I AM INCREDIBLY IMPRESSED AND PROUD. WOW → tweet link

@louszbd · 2026-07-23T10:32

GLM makes smarter tradeoffs on engineering and won the love. → tweet link

@badlogicgames · 2026-07-22T20:23

RT @NoemiTitarenco: 20 years ago, when I started writing code, the only other people writing code were terminally curious nerds like me. E… → tweet link

@badlogicgames · 2026-07-23T09:43

another question: who owns the rights to the model output? what happens when model labs change their ToS? → tweet link

@thdxr · 2026-07-23T04:34

how do you define this legally? forget the foreign part, even just domestically. who's property are the output tokens? → tweet link

@swyx · 2026-07-23T16:44

RT @aiDotEngineer: 🆕 Live now: our entire AI x Graphs Track! @emileifrem, CEO, Neo4j @yoheinakajima, Managing Partner, Untapped Capital → tweet link

@swyx · 2026-07-22T20:21

RT @yoheinakajima: 🆕 ActiveGraph: The Log is the Agent. my talk from AI Engineer is live!!! it's about @activegr… → tweet link

@FinansowyUmysl · 2026-07-22T19:05

Patrzę na LinkedIn i z tego co widzę to frontend'owcy mają bardzo, bardzo ciężko. Osoby z kilkunastoletnim doświadczeniem mają problem ze znalezieniem pracy. Najwyraźniej pierwsza grupa zawodowa, która faktycznie ucierpiała z powodu AI. → tweet link