Executive Summary
OpenAI's GPT-5.5 Instant has begun its phased rollout to all ChatGPT users, with Sam Altman calling it "a pretty big upgrade" and a signal that OpenAI is actively competing on both speed and capability. On the developer tooling front, OpenClaw and Hermes Agent are seeing rapid ecosystem expansion—OpenClaw crossed Claude Code in download volume (per TickerTrends as of April 30), while Nous Research shipped HyperFrames video generation integration and multilingual support for Hermes in under 24 hours. Coinbase's 14% workforce reduction—explicitly justified by AI enabling engineers to deliver weeks of work in days—is the most concrete corporate signal this cycle that AI productivity gains are now driving measurable headcount decisions. Meanwhile, the open-source local AI community continues to make strong arguments for consumer GPU ownership, with the NVIDIA 3090 being re-framed as the last consumer card with true NVLink for multi-GPU work.
Key Events
- GPT-5.5 Instant begins rollout to all ChatGPT users — Altman confirms it is "a pretty big upgrade" over prior models and is actively soliciting examples of use cases that required large token budgets. → link
- Coinbase cites AI productivity gains as primary justification for 14% layoffs — The company states engineers now deliver in days what previously took teams weeks, and is experimenting with "one-person teams" combining engineering, design, and PM roles. → link
- OpenClaw overtakes Claude Code in download share — Per TickerTrends data, the crossover occurred April 30, with OpenClaw's share accelerating afterward. → link
- Hermes Agent ships HyperFrames video generation skill — Native one-line install integration with HeyGen's HyperFrames allows Hermes agents to produce HTML-native videos directly. → link
- Crabbox 0.5.0 launches with authenticated WebVNC — OpenClaw's ephemeral cloud environments now support full desktop VNC with screenshot and app launch capabilities, enabling agents to reproduce and film bugs for PRs. → link
- Ollama adds Claude Desktop integration — Users can now serve Ollama Cloud models across Claude Code and Claude Cowork via
ollama launch claude-desktop. → link - "Peanut" stealth text-to-image model debuts — A new, likely indie, model appears on the Artificial Analysis Text-to-Image Leaderboard, with speculation it may be open-weights. → link
- Developer finds Codex Plus "hack": write a detailed task doc and let the agent run uninterrupted for hours — A community post details front-loading prompts into a single brief as a way to maximize Codex Plus value versus rate-limited tiers. → link
- SubQ claims true working context to 12M tokens — Carl Vellotti argues that "1M context windows are marketing numbers, not working numbers," with accuracy falling off past 200k on every frontier model. → link
- Claude Code OS workshop announced — A 3-hour hands-on workshop co-hosted with DoorDash's Hannah Stulberg focusing on team-level AI productivity systems. → link
Analysis
Acceleration in AI-driven organizational restructuring. Coinbase's explicit framing—AI is compressing team sizes and collapsing traditional role boundaries—is the clearest corporate signal in this cycle that the productivity narrative is translating into real workforce decisions. Watch for similar announcements from other scaled tech firms in coming weeks.
Developer tooling competition is intensifying. OpenClaw crossing Claude Code in downloads after a sustained update cadence (WebVNC, Google Meet, authenticated sessions) signals that tooling vendors are moving fast on agent-native features. Hermes Agent's near-daily skill drops (video, multilingual, Obsidian, Kanban) indicate the agent framework ecosystem is also highly competitive.
Context window claims face scrutiny. Multiple voices (Vellotti, LinusEkenstam) are pushing back on vendor marketing around million-token context windows, arguing working accuracy degrades significantly past 200k tokens. This could drive demand for architectures (SubQ, sub-quadratic sparse attention) that genuinely scale.
Local AI hardware advocacy is maturing. Rather than purely cost arguments, advocates like sudoingX are now making capability arguments—the 3090's NVLink being the last true multi-GPU consumer path—framing GPU ownership as infrastructure investment rather than hobby spending.
What to watch: Whether Coinbase's "one-person team" model gets replicated across other firms; whether GPT-5.5 Instant's rollout drives a measurable shift in agent framework usage patterns; and whether HyperFrames' HTML-native video generation from Hermes agents sees rapid adoption.
Tweet Feed
AI Model Releases & Research
@sama · 2026-05-05T17:33
5.5 instant comes to ChatGPT today! imo it is a pretty big upgrade, i really like using it. → tweet
@sama · 2026-05-05T18:04
RT @michpokrass: we shipped gpt-5.5 instant today to chat; it's rolling out over the next couple days to everyone. for this model, we focus… → tweet
@sama · 2026-05-05T18:04
i would like to talk to people who have built amazing things with 5.5 that weren't possible with earlier models. i am especially interested in examples that took ludicrous token budgets. thanks. → tweet
@sama · 2026-05-05T00:51
pretty excited for voice models to get great its interesting to watch how people are already starting to change the way they interface with AI → tweet
@sama · 2026-05-05T14:27
we have very efficient models, especially for their capability level happy codexing → tweet
@Ex0byt · 2026-05-04T21:20
"Peanut" - is a brand-new, stealth text-to-image model which just debuted today on the Artificial Analysis Text-to-Image Leaderboard. I have a hunch this one is from an indie shoppe and will be open-weights.. → tweet
@LinusEkenstam · 2026-05-05T14:32
You spent months engineering around context collapse. Better prompts. Smarter chunking. Careful retrieval. I've seen this before. Teams doing a lot of work, just so that one update later, all of that work gets obsolete. This is one such update. Sub-quadratic sparse-attention architecture, thats a mouthful, but 12 million token context window!? → tweet
@juliarturc · 2026-05-05T01:33
And this is why we need world models. Made with Veo 3. What in the transformers is this? → tweet
OpenAI / Codex CLI
@sudoingX · 2026-05-05T07:28
few days into codex plus and i think i found the hack. nobody is talking about it and the value sitting in this subscription is wild. the hack: do not prompt the agent. write a single detailed task doc with every requirement laid out plus the final vision of what you are building, then fire codex cli with one line, accomplish this and test until done. → tweet
@TheAhmadOsman · 2026-05-05T00:04
LLMs 101 Tokens are not words → tweet
@nummanali · 2026-05-04T19:29
DeepSeek V4 in Codex through Responses API Works in CLI and the Codex App. Top tip, this works for any model through @OpenRouter → tweet
Developer Tools & Frameworks
@steipete · 2026-05-05T08:28
🤖 Kept hitting @github rate limits across my agents. Shipped two things: – RepoBar got a JUICE METER – gitcrawl is now also a drop-in gh cache → symlink it as gh, reads served from local SQLite → tweet
@steipete · 2026-05-05T08:09
gog 0.16 is out. Google Workspace CLI for humans and agents. Lossless raw API output, sanitized Gmail reads, safer command profiles, Drive inventory, Docs tabs, Sheets tables, Gmail filter export, and official Docker images. → tweet
@steipete · 2026-05-05T02:15
Crabbox 0.5.0 is live 🦀 🖥️ Desktop/browser leases 🧑💻 VNC + authenticated WebVNC 🪟 AWS Windows + WSL2 📸 Screenshots + app launch. Remote CI boxes, now suspiciously usable. → tweet
@steipete · 2026-05-05T16:18
I added Googe Meet support to OpenClaw and now Molty is eager to join every meeting. → tweet
@steipete · 2026-05-05T06:58
We can now reproduce issues directly in empheral crabboxes with WebVNC (Linux/Windows/macOS). Agents set up the exact state to test + fix and post videos on the PR. Working hard to level up our QA. → tweet
@steipete · 2026-05-05T16:55
I asked Molty to review my PR and it made a song. → tweet
@badlogicgames · 2026-05-05T14:25
agentOS by @rivet_dev is nifty, and not just because it's built on https://t.co/oUoqqL9hAp. check it out! → tweet
@ollama · 2026-05-04T23:36
🤯 Ollama now supports Claude Desktop via Claude's built-in third party inference. ollama launch claude-desktop. This allows all models from Ollama's Cloud to be used across Claude Cowork and Claude Code from the Claude Desktop app. → tweet
@steipete · 2026-05-05T11:38
RT @ashleywolf: We're hosting OpenClaw 🦞 After Hours at the GitHub SF office on June 3. It'll be a claw-some line up of speakers from OpenC… → tweet
@steipete · 2026-05-05T12:12
RT @tickerplus: Codex has overtaken Claude Code in downloads. TickerTrends shows the crossover on April 30, followed by accelerating share… → tweet
@steipete · 2026-05-05T12:12
RT @jlehman_: It really is. Similar vibe to Opus of yore but without the constant made-up shit. → tweet
@steipete · 2026-05-05T07:02
RT @manan: I actually didn't fully understand how effective this was till I submitted a bug. Clawsweeper looked at it, validated it, said n… → tweet
@steipete · 2026-05-05T07:04
RT @GeoffreyHuntley: no longer using opus for code. very happy with 5.4/5.5. → tweet
@steipete · 2026-05-05T17:39
RT @mronge: @steipete the latest OpenClaw release is fantastic! Fixes all the issues I've having the past week. Thank you! → tweet
@badlogicgames · 2026-05-05T14:25
RT @steipete: It's been quite a week. Good stuff is coming though. I hired a team! → tweet
Hermes Agent / Nous Research
@Teknium · 2026-05-05T15:24
Hermes now speaks your language. Hermes 现在会说你的语言。Hermes があなたの言語を話すようになりました。Hermes spricht jetzt deine Sprache。Hermes ahora habla tu idioma. → tweet
@Teknium · 2026-05-05T16:37
RT @HeyGen: Hermes Agent now has @Hyperframes_ skill, natively. a close collab with @NousResearch. One-line install $ hermes skills instal… → tweet
@Teknium · 2026-05-05T17:08
Welcome to Hermes Agent crew! Reach out if you have any suggestions or issues! → tweet
@Teknium · 2026-05-05T12:13
Quickstart guide got an upgrade thanks to @tonbistudio 🔥🔥🔥 → tweet
@Teknium · 2026-05-05T11:46
A great installation, getting started tutorial by the legendary Tonbi, check it out! → tweet
@Teknium · 2026-05-05T12:39
RT @shmidtqq: hermes users be like: "it's a chatbot" meanwhile the agent has: - persistent memory - filesystem rollback - session branchi… → tweet
@Teknium · 2026-05-05T23:19
RT @ivanfioravanti: Hermes Agent + Obsidian is match in heaven! I'm dumping all info, meetings, ideas to hermes and I get back well formatt… → tweet
@Teknium · 2026-05-04T23:19
Welcome to Hermes Gang! → tweet
@Teknium · 2026-05-05T01:50
RT @gkisokay: My Hermes agents get smarter every day, and the models are not the reason. It's one upstream research agent feeding the enti… → tweet
@Teknium · 2026-05-05T16:57
RT @realsigridjin: @Teknium @NousResearch now cooking the posters for the first hermes community meetup @ seoul, happening this saturday… → tweet
@Teknium · 2026-05-04T23:08
RT @ruffy0369: Just shipped a meta-compiler for @NousResearch hermes. It reads papers and turns them into optimized triton kernels for age… → tweet
@kunchenguid · 2026-05-05T18:38
one of the things I'm doing now that weren't possible without AI is OPINIONS.md. I get a ton of value and believe everyone should create one. I manage mine with Hermes agent from @NousResearch and gpt 5.5 → tweet
Local AI & GPU
@sudoingX · 2026-05-05T16:00
you cannot describe the taste of free thinking until you own the machine that runs it. the day the model moves from someone else's cloud to your own desk, the way you think changes. every kid in this decade should grow up with their own gpu and their own model. → tweet
@sudoingX · 2026-05-05T09:29
the last consumer gpu that lets you build real multi gpu local infrastructure is a card from 2020. ampere was the cutoff. 3090 has the bridge, full p2p memory pooling across two cards over 112 gb/s nvlink. run a 70b at full precision across 48gb pooled, train across multi cards, do real multi-gpu work without datacenter pricing. → tweet
@sudoingX · 2026-05-05T17:13
many assume fine tunes are slower than the base model. ran the test myself tonight, both vanilla qwen 3.6-27b and carnice-v2 trained on hermes agent traces, same flags, same kv cache, on my rog 5090 mobile 24gb vram. the numbers came back tied within 1% on every metric. → tweet
@sudoingX · 2026-05-05T14:48
a question keeps hitting my mind. does same base SFT for harness beat vanilla qwen 3.6-27b on hermes agent agentic loops? to find out i loaded carnice-v2 by kaios, qwen 3.6-27b tuned specifically on hermes agent traces. been wanting to run this exact lineup. results coming next anon. → tweet
@sudoingX · 2026-05-05T15:26
carnice-v2 is up now on my rog 5090 mobile 24gb tier, llama-server holding 21gb at 262k context, hermes agent on the other pane, nvtop watching the line go up. hell yeah tonight is gonna be great. if you are running with me on 24gb vram, here is the exact stack... → tweet
@sudoingX · 2026-05-05T14:15
just stopped to give hermes agent a task lol → tweet
@TheAhmadOsman · 2026-05-04T20:47
PRO TIP: Using local LLMs? Give them a web stack. My setup: SearXNG for candidate source discovery, Firecrawl for known-URL scraping, Camofox for browser fallback. Search → Extract → Interact. → tweet
@TheAhmadOsman · 2026-05-05T15:42
Please stop saying that the DGX Spark or any Unified Memory machine competes with GPUs in any meaningful way. It's misleading. Speak of the pros and cons of each, but stop misinforming your audience for whatever reason you're doing it for. → tweet
@tinygrad · 2026-05-05T17:00
TIL why the Blackwell box is out of stock in Singapore. It's too powerful! Get yours before the regime clamps down further. → tweet
@tinygrad · 2026-05-05T13:55
A reminder that we have two tinyboxes you can buy today. Bring the power of AI into your home or office. → tweet
@FrameworkPuter · 2026-05-05T13:55
RT @svpino: The Framework 13 Pro is the best PC laptop I've ever tried. For the first time in a long time, I'm excited about computers aga… → tweet
@nummanali · 2026-05-05T17:33
MTP is increasing speeds of on device inference by 2x - 3x. That's going from 20tps to 60tps! Gemma 4 models now have official support. → tweet
Industry & Business
@FinansowyUmysl · 2026-05-05T11:31
Coinbase właśnie ogłosił zwolnienie 14% osób, ale ciekawsze są nie tyle zwolnienia, co uzasadnienie: "(...) W ciągu ostatniego roku obserwowałem, jak inżynierowie wykorzystują sztuczną inteligencję, aby w ciągu kilku dni dostarczyć to, co kiedyś zajmowało zespołowi tygodnie. Zespoły nietechniczne dostarczają teraz kod produkcyjny, a wiele naszych procesów jest automatyzowanych. (...) Będziemy również eksperymentować z mniejszymi rozmiarami zespołów, w tym z „zespołami jednoosobowymi" składającymi się z inżynierów, projektantów i menedżerów produktu pełniących jednocześnie jedną rolę." → tweet
@FinansowyUmysl · 2026-05-05T18:36
RT @FinansowyUmysl: Coinbase właśnie ogłosił zwolnienie 14% osób, ale ciekawsze są nie tyle zwolnienia, co uzasadnienie... → tweet
@kunchenguid · 2026-05-04T23:15
what does it feel like to quit your job and go solo? it's been just a month since I quit my big tech career as an L8 engineer. i wrote down my experience of this past month in the most transparent way possible. hope it's helpful to fellow builders! → tweet
@steipete · 2026-05-05T10:39
It's been quite a week. Good stuff is coming though. I hired a team! → tweet
@thdxr · 2026-05-05T04:36
RT @jayair: OpenCode Go is now our second breakout product. Closing in on 1T tokens per day. Shoutout to Frank and the team → tweet
@victormustar · 2026-05-05T10:05
RT @stevibe: We have so many incredible open-source models in 2026, something we couldn't have imagined just a year ago. → tweet
AI Productivity & Workflow
@carlvellotti · 2026-05-05T18:46
AI productivity gains have a natural limit. You really can't move faster than your team. Even if you have the PERFECT AI setup... You'll still get blocked in all kinds of ways. BUT what if you could unlock those productivity gains at a TEAM level? When you get it right: Customer call themes auto-synthesize into a weekly Slack digest. Meeting notes auto-extract decisions, owners, and follow-ups. Hannah Stulberg is building this out right now at DoorDash. How to Build Your Claude Code OS — an action-packed, extremely hands-on 3-hour workshop. 17 realistic demos, a fully built out example Team OS. → tweet
@carlvellotti · 2026-05-05T14:56
the 1M context windows the labs sell us are marketing numbers, not working numbers. accuracy falls off a cliff past 200k on EVERY frontier model. subq is the first thing i've used that actually holds context… all the way to 12M tokens. → tweet
@LinusEkenstam · 2026-05-05T15:15
Claude will become the OS. Eating up software layers one repository at the time. → tweet
@LinusEkenstam · 2026-05-05T18:56
Me using Codex circa 2021. We've come a long way since → tweet
@iamdevloper · 2026-05-05T16:26
managers showing back up as ICs → tweet
@iamdevloper · 2026-05-05T16:25
LLMs haven't made points around all this any easier, they've just allowed agents just ship more bugs in parallel. → tweet
@jezell · 2026-05-05T16:59
Even Anthropic can't get people to merge their vibe coded spaghetti. → tweet
@MengTo · 2026-05-05T15:26
GPT 5.5 is surprisingly good at recreating UIs from a URL, DESIGN.md or designs made in Images 2.0. It struggles with animations and WebGL, but you can work around that by providing the skills or HTML. → tweet
@RydMike · 2026-05-05T08:02
RT @diegohaz: I gave the same task to GPT-5.5 xhigh and Opus 4.7 max. GPT took ~30 minutes. Opus took ~2 hours. Then I asked them t… → tweet
@RydMike · 2026-05-05T14:46
RT @spydon: Don't miss out on the most fun Flutter conference this autumn! → tweet
Programming & Open Source
@badlogicgames · 2026-05-05T07:53
recommended viewing because @karpathy → tweet
@badlogicgames · 2026-05-04T20:27
recommended reading! → tweet
@badlogicgames · 2026-05-05T07:58
RT @davis7: This is very late, but I'm finally done with my 5.5 vid - use low reasoning - the name sucks - it's fast - best code I've ever… → tweet
@badlogicgames · 2026-05-04T20:30
RT @cmpatino_: Introducing nanowhale 🐳! A tiny DeepSeek model fully pretrained by an agent. Inspired by @karpathy's nanochat, we gave ml-i… → tweet
@badlogicgames · 2026-05-04T20:26
RT @antirez: [blog post] Redis array: short story of a long development process => https://t.co/Q5paOZH2Vz → tweet
@badlogicgames · 2026-05-05T15:00
had the good fortune to talk to @davidond on his podcast. → tweet
@badlogicgames · 2026-05-04T21:30
RT @_lopopolo: @badlogicgames @smcllns @dexhorthy @GeoffreyHuntley @mattpocockuk This is also why things like SDK design are hard for agent… → tweet
@badlogicgames · 2026-05-05T08:33
RT @steipete: We can now reproduce issues directly in empheral crabboxes with WebVNC (Linux/Windows/macOS). Agents set up the exact state… → tweet
@jsuarez · 2026-05-05T17:45
PufferLib RL dev with Joseph Suarez → tweet
@jsuarez · 2026-05-05T12:43
Heading to NYC tomorrow. Will be there Thursday + Friday. Puffer builds the highest performance RL tools and we do custom sims professionally. DM/email - I still have some availability to present our latest breakthroughs in person → tweet
@ASalvadorini · 2026-05-05T07:03
RT @birjuvachhani: Got this @dart_lang plugin update today! I guess we will be having primary constructors feature in this #GoogleIO?! Alth… → tweet
@RydMike · 2026-05-05T17:11
RT @remi_rousselet: #Riverpod folks: There's an open proposal for a possible new syntax without using code-generation. → tweet