Executive Summary
OpenAI launched GPT-5.5, accompanied by new Codex workspace agent features and ChatGPT for Clinicians, marking a significant model and product release day. Qwen 3.6 27B Dense arrived as a major open-source model, achieving impressive benchmarks on consumer hardware (39.82 tok/s on a single RTX 3090) and immediately becoming the flagship local AI model. The Hermes Agent ecosystem continued its rapid expansion with support for Qwen 3.6, Kimi K2.6, Xiaomi MiMo-V2.5, and OpenAI image generation. Anthropic published a post-mortem acknowledging Claude Code quality regressions caused by harness bugs, while OpenClaw-style agents reached Codex business accounts.
Key Events
- OpenAI releases GPT-5.5 — described as "a new class of intelligence" with token efficiency, low latency, and the ability to complete challenging tasks with minimal micromanagement. → link
- Qwen 3.6 27B Dense launches — open-source model available on Ollama, achieving 39.82 tok/s on a single RTX 3090 (13% faster than Qwen 3.5 baseline), with strong reasoning even on greetings. → link
- Anthropic publishes Claude Code post-mortem — coding regressions traced to harness bugs (stale sessions losing thinking, bad system prompts), not model changes. Fix in v2.1.116+. → link
- Hermes Agent ecosystem expands massively — adds Qwen 3.6, Kimi K2.6, Xiaomi MiMo-V2.5, OpenAI image generation OAuth; $26K hackathon ongoing; 5,350+ members in 29 days. → link
- OpenAI ships Codex workspace agents — cloud-hosted agents for teams with Slack integration, recurring tasks, and tool hooks; OpenClaw-style agents now available for business accounts. → link
- OpenAI releases Images 2.0 — Sam Altman describes it crossing "an important qualitative threshold"; integrated into Bearly AI and Hermes Agent. → link
- Tinygrad claims 420 tok/sec on AMD 7900XTX — custom driver beating a 3090 megakernel; AMD open-sources SQTT profiling format, signaling deep open-source commitment. → link
- Framework Laptop 13 Pro sells out — first six batches gone; Ubuntu configurations outselling Windows configurations. → link
- ChatGPT for Clinicians launched — free version of ChatGPT designed for healthcare professionals. → link
- OpenAI drops PII detection model — 1.5B params (50M active), labels all sensitive data in one pass. → link
Analysis
Local AI ownership is becoming a cultural movement, not just a technical preference. Multiple influencers (sudoingX, TheAhmadOsman) are aggressively pushing the narrative that renting AI by the token is a losing strategy, advocating for consumer GPU ownership with quotes like "I don't want my best ideas going through a meter." The arrival of Qwen 3.6 27B Dense—running at near-frontier quality on a $700 used RTX 3090—gives this movement serious ammunition.
The agent layer is the new competitive battleground. OpenAI's Codex workspace agents, Hermes Agent's explosive community growth, Claude Code's harness regression, and OpenCode's rising popularity all point to agents as the primary vector of competition. The critical insight from Claude Code's post-mortem is that harness quality, not model quality, was the regression driver—"the model didn't change, the harness did."
Open-source model velocity is accelerating past proprietary release cadences. In the last 30 days, three frontier-adjacent models shipped with downloadable weights (Nvidia Nemotron, Google Gemma 4, Qwen 3.6). Kimi K2.6 is being called "surprisingly good" by builders who accidentally started using it. Xiaomi's MiMo models are now in production agent frameworks.
What to watch: (1) Whether GPT-5.5's "new class of intelligence" claim translates to measurable benchmarks or remains qualitative; (2) AMD consumer GPU benchmarks for Qwen 3.6 — a data gap the community is actively crowdsourcing; (3) Codex workspace agents expanding to personal accounts; (4) Whether Claude Code's harness fix restores user trust or if the exodus to open alternatives (OpenCode, Hermes) continues; (5) Tinygrad's AMD custom driver claims — if verified, this reshapes the consumer AI hardware landscape.
Tweet Feed
GPT-5.5 Release & OpenAI
@sama · 2026-04-23T18:13
GPT-5.5 is here! We hope it's useful to you. I personally like it. → tweet link
@sama · 2026-04-23T18:18
- We believe in iterative deployment; although GPT-5.5 is already a smart model, we expect rapid improvements. Iterative deployment is a big part of our safety strategy; we believe the world will be best equipped to win at the team sport of AI resilience this way. 2. We believe in democratization. We want people to be able to use lots of AI; we aim to have the most efficient models, the most efficient inference stack, and the most compute. We want our users to have access to the best technology and for everyone to have equal opportunity. We have been tracking cybersecurity as a preparedness category for a long time, and have built mitigations we believe in that enable us to make capable models broadly available. 3. We love you and we want you to win. We want to be a platform for every company, scientist, entrepreneur, and person. → tweet link
@gdb · 2026-04-23T18:24
GPT-5.5 is a new class of intelligence. This intelligence makes it very intuitive to use; it can complete very challenging tasks with little micromanagement. It is also very token efficient, and we can run it with low latency and at scale. Feels like a real step toward a new way of doing computer work. → tweet link
@sama · 2026-04-23T18:13
Also, a ton of new Codex features coming soon! Fun little bundle w/the new model. → tweet link
@sama · 2026-04-23T16:18
Images 2.0 really got over some important qualitative threshold for me that I didn't know existed. → tweet link
Codex & OpenAI Platform
@gdb · 2026-04-22T18:41
Build workspace agents for your team, on top of a cloud-hosted Codex harness. Hook them up to tools, give them recurring tasks, and talk to them from surfaces like Slack. Easier than ever to bring the power of agents to your computer work. → tweet link
@nummanali · 2026-04-22T18:49
OpenClawesque Agents are now in Codex. Waiting for it to come to personal accounts, currently only business :( → tweet link
@gdb · 2026-04-22T21:27
ChatGPT plugin now available for Google Sheets: → tweet link
@gdb · 2026-04-23T02:46
Introducing ChatGPT for Clinicians: → tweet link
@jezell · 2026-04-23T16:37
Surprised it took this long to make it into the Responses API, but looks like video input is finally here. → tweet link
@Ex0byt · 2026-04-23T00:51
Someone at OpenAI is clearly a follower. lol.. OpenAI dropped a tiny PII detection model earlier today. 1.5B params (50M active), labels all sensitive data in one pass and masks it out. → tweet link
Qwen 3.6 27B Dense
@sudoingX · 2026-04-22T19:22
this was supposed to be a normal evening, then i saw on the timeline that qwen 3.6 27b dense q4 weights from unsloth are live and i could not sit still. compiled llama.cpp with cuda on the single rtx 3090 at 2am from bangkok... first pass said "yo" to the model as a warmup. it ran a six step thinking chain to formulate "yo what's up how can i help you today". full reasoning visible... 39.82 tokens per second on the first real generation. march baseline on this exact hardware was 35.3 flat... that is a 13 percent speed bump. → tweet link
@sudoingX · 2026-04-23T17:41
before i touch any turbo or quant tricks on the new qwen 3.6-27b dense, i ran a full context sweep on a single 3090. same flags as the march qwen 3.5 baseline. same hardware. the architecture is inherited. exact same vram footprint as 3.5 at every context: 16 gigs at 4k, 18 at 128k, 21 at 262k, 23 at the 376k vram wall. and 13.7 percent faster on identical config. 40.13 tok/s vs 35.30. → tweet link
@sudoingX · 2026-04-23T17:55
dude! the new qwen 3.6-27b dense is hammering my single 3090 at 100% gpu utilization. the spiky pattern on nvtop is the hermes agent autonomously thinking, calling tools, reading results, thinking again. this model is so cool to talk to. waits for tool outputs, reads them, selfcorrects, keeps going. no stalls, no loops, no hand holding. → tweet link
@ollama · 2026-04-22T21:33
Qwen 3.6 27B model is available on Ollama! Use it with all the integrations in Ollama or chat with the model. → tweet link
@TheAhmadOsman · 2026-04-22T22:34
This prediction was right, and on a single 5090 not even RTX PRO 6000. Qwen 3.6 27B is the at-home flagship model right now → tweet link
@TheAhmadOsman · 2026-04-23T02:45
Qwen 3.6 27B means the permanent underclass thing has been canceled btw → tweet link
Hermes Agent Ecosystem
@Teknium · 2026-04-23T04:34
You can now use OpenAI OAuth to generate images with Hermes Agent! Update Hermes and use the 'hermes tools' command to configure the image generation tool to use OpenAI! Reminder: we have a $26,000 online creativity with Hermes Agent hackathon ongoing. → tweet link
@Teknium · 2026-04-22T23:20
Xiaomi's MiMo-V2.5 and MiMo-V2.5-Pro are both now available in Hermes Agent through Nous Portal and OpenRouter! Just
hermes update! → tweet link
@Teknium · 2026-04-23T03:08
If you have had any troubles with the GLM subscription, Hermes is now officially supported and whitelisted! → tweet link
@Teknium · 2026-04-23T22:57
Another amazing tutorial / deep dive by @tonbistudio on Hermes Agent - this one covering the power and value of WebHooks. → tweet link
@TheAhmadOsman · 2026-04-23T08:26
This is how Hermes agent creates skills and develops memories btw → tweet link
Claude Code & Anthropic
@nummanali · 2026-04-23T18:02
TLDR: March 4th > April 7th - Default reasoning was medium over high. March 26th > April 10th - Sessions stale over one hour had thinking removed due to bug. April 16th > April 20th - Bad system prompts caused coding regression. Upgrade to v2.1.116+ to be on fixed Claude Code → tweet link
@badlogicgames · 2026-04-23T17:56
recommended reading. cool they are fixing things. but it's also a reason i switched away from CC. no control over the harness means having to wait for them to fix things. the model didn't change. the harness did. → tweet link
@TrungTPhan · 2026-04-22T18:44
Anthropic: "Claude Mythos is the most dangerous software in the history of the most dangerous software." Anthropic's security for Claude Mythos: → tweet link
Local AI & Self-Hosted Models
@sudoingX · 2026-04-23T17:13
the deeper i get into this, the less i want intelligence i have to rent by the sentence. i don't want my best ideas going through a meter. i don't want my curiosity rate limited. i don't want my workflow shaped by somebody else's pricing page. my thinking i need to own. → tweet link
@sudoingX · 2026-04-23T09:34
in the last thirty days open source shipped nvidia nemotron, google gemma 4, and qwen 3.6 27b dense. three frontier adjacent models you can download and run on consumer hardware right now. while you are reading this, a corporate salesman somewhere is getting paid to tell you that local ai is not there yet. → tweet link
@TheAhmadOsman · 2026-04-23T15:40
Alternatives to Anthropic and OpenAI you can self-host so your data stays yours: Kimi K2.6, GLM 5.1, MiniMax M2.7, Qwen 3.5 / 3.6. Plug them into your favorite agent. Stop donating your business data. → tweet link
@TheAhmadOsman · 2026-04-23T16:49
Gentle reminder that all you need to start with local AI is: 2x RTX 3090s, Qwen 3.6 27B, Your favorite agent, Self-hosted SearXNG for web access. And you got yourself Opus 4.5 at home → tweet link
@sudoingX · 2026-04-23T09:41
your gpu does not matter if you have not written one evaluation for your actual work. → tweet link
Consumer GPU Benchmarking & Kernel Optimization
@sudoingX · 2026-04-23T08:53
this guy just cracked 134 tok/s on qwen 3.5-27b dense and 73 on new qwen 3.6-27b on a single 3090. open source moves at godspeed in 2026. weights ship in the evening, dynamic ggufs land by midnight, fused kernel + speculative decoding stack runs the new model 12 hours after release. → tweet link
@sudoingX · 2026-04-23T09:07
hey if you are running new qwen 3.6 27b dense on an rtx 4090... @Punch_Taylor ran my exact flags on 4090 wsl2 ubuntu cuda 13.2, three warm runs on q4_k_m. average landed at 43.1 tok/s, 8.3 percent above my 3090 baseline of 39.82. that delta tracks the memory bandwidth gap almost perfectly. → tweet link
@sudoingX · 2026-04-23T10:14
calling all amd users my dms and replies are stacking with amd owners asking how their card runs qwen 3.6 27b dense. i do not own any red machines. i cannot share data i do not have. so i am sharing this asking if you run amd hardware, any tier, drop your numbers. → tweet link
Tinygrad / AMD
@tinygrad · 2026-04-23T04:58
This megakernel is using a 3090. Stock tinygrad beats this (420 tok/sec!) using a cheaper 7900XTX. With our custom driver AMD hardware can really shine. → tweet link
@tinygrad · 2026-04-23T05:09
Giving us the MI300X boxes marked the turning point. Since then, AMD open sourced their SQTT (low level profiling) format, which is a commitment to open source beyond what I expected. If good decisions like this keep being made, this is just the beginning. Hyped for RDNA5. → tweet link
Hardware: Framework & AMD
@TheAhmadOsman · 2026-04-23T18:09
New hardware from Framework / AMD: Max+ 395 - 128GB Desktop PC - 16-core/32-thread, 64MB L3 Cache - 3.0GHz, up to 5.1GHz max boost - 128GB LPDDR5x-8000 - Radeon 8060S Graphics - Wi-Fi 7, 5Gbit Ethernet → tweet link
@FrameworkPuter · 2026-04-23T15:55
Framework Laptop 13 Pro is selling far above our forecast, and we've sold out of the first six batches already. Also nice validation of our approach, the Ubuntu configurations are outselling the Windows ones! → tweet link
@alexinexxx · 2026-04-23T18:01
the year of linux on desktop has finally arrived → tweet link
Developer Tools & Open Source
@levelsio · 2026-04-23T13:57
I open sourced my first Chrome extension 🚀 SuperLevels... I vibe coded it to replace all my Chrome extensions that are increasingly being bought up by spyware and malware companies who sell your data or worse hack your accounts... Features: Tab Cleaner, Cookie Editor, Redirect Tracer, Dark Mode, X Dim Mode, JS Toggle, GDPR Cookie Consent Dismisser, Live CSS Editor, YouTube Unhook, Music Recognizer, Picture-in-Picture, Google Maps Links, View Image, JSON Formatter → tweet link
@nummanali · 2026-04-23T09:39
Hello Web/Node/Bun/TS/JS Devs - QQ! Would you use a TypeScript ML library built entirely for Mac MLX? → tweet link
@Ex0byt · 2026-04-23T15:13
We now have amazing local models (and more will come). We need better harnesses. Stop listening to the blue-pill shills. Read the paper and audit your own harnesses. → tweet link
@jsuarez · 2026-04-23T13:23
Raylib rocks - we even used it for our new experiment dashboard! If you're using raylib at your company, see if you can get them to pitch in a couple hundred bucks a month to support the project → tweet link
SWE-Bench & Model Analysis
@badlogicgames · 2026-04-23T17:10 (RT @SrihariSriraman)
I analyzed 730 SWE-Bench Pro trajectories each for Sonnet 4.5 and GPT-5 and turned them into "trajectory shapes": when... → tweet link
@badlogicgames · 2026-04-23T17:11 (RT @SrihariSriraman)
I wanted to check this on newer models, but no SWE-bench Pro trajectories exist for Opus 4.6 / GPT-5.4. So I pulled @... → tweet link
DeepSeek Kernels
@alexinexxx · 2026-04-23T15:22 (RT @eliebakouch)
new deepseek kernels released, both forward and backward for Engram > some of them have already been used in internal tra... → tweet link
AI-Generated Content & Web4 Concerns
@nummanali · 2026-04-23T08:47
Cool in principle but with such easy access there will be a wave of AI generated repos, websites, reports, papers... We're skipping Web 3 and going straight to Web 4 - which is currently looking a little messy. I wonder if search engines and curators will deploy standardised analysers to give a scoring on AI Generated content → tweet link
World Models & Real-Time Generation
@LinusEkenstam · 2026-04-23T13:53
"Every pixel will be generated" — Jensen Huang. World models have been streaming for some time but they have a temporal cut-off at very short session times... This approach is different, and heavily optimized version of LTX-2 has been used to make it possible to live stream directly from a GPU in the cloud at 24fps into your browser. You're essentially seeing 1080p diffusion happening 24 times per second, thus creating an endless video stream. → tweet link
@juliarturc · 2026-04-22T21:36
Currently working on a video about world models. Let me know your questions! 👇 → tweet link
Robotics
@LinusEkenstam · 2026-04-23T13:43
The general public have very little knowledge into just how fast robotics is advancing... Hardware components in a robot are essentially e-mobility + cameras + compute... Now the brain part of a robot, what makes it actually become a part of life, work or leisure is looking more and more to be agentic... A humanoid will only get better after you purchase one. Software will be moving as fast as your average AI chat bot. → tweet link
@TrungTPhan · 2026-04-23T16:39
We need to see Sony's expert AI ping pong robot (Ace) play Jensen... → tweet link
AI Video Tools
@LinusEkenstam · 2026-04-23T17:28
Create videos from just texting. This is Cappy. Your new always on video editor. When the UI disappears, you start focusing on what matters more. The story you are trying to tell. → tweet link
Multi-Agent Simulation
@kunchenguid · 2026-04-22T19:59
lol remember this org chart meme? I just created a full simulation for all of them with agents, and the results blew my mind! the simulation asked each organization to build and ship a web spreadsheet → tweet link
@kunchenguid · 2026-04-23T14:55
this blew up so adding a quick clarification - despite the agents proved to be highly human-like in the observations, this is just a simulation meant to explore the effects of multi-agent topologies. this is NOT meant to reflect the capabilities of those real companies! :) → tweet link
AI-Assisted Development Workflows
@badlogicgames · 2026-04-23T17:16
my workflow is mostly: prompt template with injected gh issue url and instructions... i confirm analysis/plan... tell model to implement... check results, steer if necessary... if all good, another prompt template is used to wrap up, i.e. changelog, docs, commit, push, comment on issue, close issue → tweet link
@badlogicgames · 2026-04-23T15:11
4yo used my little TTS desktop app + pi, tells the agent what to build, what to change. more dinosaurs to select from, now working on king kong as final boss. i love this so much. slop doesn't matter. early agency does. → tweet link
AI for Trading Strategies
@FinansowyUmysl · 2026-04-23T07:09
...wczoraj znalazłem chwilę czasu i pomyślałem, że wykorzystam nowe skille z claude code (superpowers) i zlecę im znalezienie lepszych parametrów... i jestem zaskoczony, jest kilka ustawień dla których moja strategia byłaby lepsza... nawet wzrost CAGR z ~35% do ~50% przy tym samym DD → tweet link
Reinforcement Learning
@jsuarez · 2026-04-23T15:51
Reinforcement Learning dev with Joseph Suarez → tweet link
Security
@jezell · 2026-04-23T04:42
Maybe Apple put a hardware guy in charge because software is cooked. The future of household software is services. No one needs a new phone. They need a robot to do the dishes and clean the house. → tweet link
@jezell · 2026-04-23T20:42
Transparent secret proxies are the new http only cookies: → tweet link
@steipete · 2026-04-23T17:11
My favorite security advisories. → tweet link
Google Cloud Next / Dart
@ASalvadorini · 2026-04-23T07:10 (RT @sgnagnarella)
Dart for Cloud Functions - A massive win for the Flutter community! You can now write serverless backends in Dart. → tweet link