Executive Summary
The past 24 hours in AI/tech have been dominated by GPT-5.5's rollout, with Sam Altman personally praising its "switched on" capabilities while developers report mixed experiences with its coding performance. Local AI inference continues maturing—@sudoingX demonstrated a 27B model (Carnice v2 on Hermes Agent) completing complex multi-step tasks including self-benchmarking and verification on consumer hardware. The ecosystem is also seeing significant tooling activity: OpenClaw released multiple updates, inference engine comparisons went viral, and debates around SaaS displacement by AI intensified as @levelsio documented replacing multiple services with custom AI solutions. Hardware discussions centered on NVIDIA's 5090 mobile performance and new DGX Spark workstation capabilities.
Key Events
-
GPT-5.5 Instant launches — Sam Altman confirms "ChatGPT feels very switched on" and praises the new instant model's speed, intelligence, and memory; @gdb reports major upgrade rolling out → tweet
-
Comprehensive inference engine breakdown goes viral — @TheAhmadOsman published detailed cheat sheet covering llama.cpp, MLX, ExLlamaV2/V3, vLLM, SGLang, and TensorRT-LLM with use-case recommendations → tweet
-
OpenClaw ecosystem expands — @steipete shipped iMsg 0.6/0.7 with private API bridge, CodexBar 0.24 with new Windsurf/Codebuff/DeepSeek providers, and openclaw/fs-safe filesystem safety primitive → tweet
-
Local AI agent demonstrates full self-benchmark workflow — @sudoingX showed Carnice v2 27B (on Hermes Agent traces) completing hardware detection, git probing, self-benchmarking via curl, markdown report generation, and verification loop on 5090 mobile → tweet
-
Hermes Agent adds plugin extensibility — @Teknium announced expanded surface area for plugins including LLM inference providers and gateway channels → tweet
-
Local AI cost argument intensifies — Multiple posts from @sudoingX and @levelsio arguing that 90%+ of AI startup API costs could run locally on consumer GPUs → tweet
-
MRC networking protocol released — @gdb announced Multipath Reliable Connection for large AI training clusters → tweet
Analysis
Escalation Patterns: - Developer tooling wars intensifying with OpenClaw vs. Cursor vs. Codex vs. traditional IDEs receiving significant attention and complaints - Local AI narrative shifting from "experimental" to "production-viable" for many workloads, though @sudoingX's testing shows hype around optimization techniques (turboquant) often exceeds reality - SaaS displacement discussion accelerating with concrete examples of AI replacing services
De-escalation/Frustration Signals: - @badlogicgames expressing frustration with agents ("garbage", "jfc imma stop using agents") - VS Code autocomplete issues reported ("basically useless now") - Cursor team accused of ignoring developer outreach - Criticism of frontier models' "retarded" frontend/design execution despite strong backend capabilities
What to Watch: - GPT-5.5's practical performance in production workflows vs. Altman praise - Hermes Agent ecosystem growth (1,400+ stars on HUD Web UI) - Whether local AI claims translate to real enterprise adoption - xAI/Anthropic compute partnership implications - Inference engine benchmarks comparing real-world performance
Tweet Feed
AI Model Releases & Performance
@sama · 2026-05-06T01:00
ChatGPT feels very 'switched on' now → tweet
@sama · 2026-05-05T20:17
the new instant model in chatgpt is so good damn if you have been thinking-model-only for awhile, give it a try! → tweet
@sama · 2026-05-05T20:18
in particular, the combination of improvements to speed, intelligence, personality, and great memory/personalization feels like a more-than-sum-of-the-parts thing when it all hits together → tweet
@gdb · 2026-05-05T20:24
Major ChatGPT upgrade rolling out now, in the form of GPT-5.5 Instant: → tweet
@TheAhmadOsman · 2026-05-06T07:17
Current SoTA LLMs Ranking GPT 5.5 > Kimi K2.6 > GLM 5.1 > Qwen 3.6 397B > MiniMax M2.7 And yes, Opus 4.7 is slop, I left Claude models out intentionally → tweet
@TheAhmadOsman · 2026-05-06T07:03
He could've just said Buy a GPU lmao → tweet
@TheAhmadOsman · 2026-05-06T03:35
The 3 most consequential open weight releases - Llama 3 - Qwen 2.5 - DeepSeek R1 Opensource AI wouldn't be here today without those 3 Forever grateful → tweet
@TheAhmadOsman · 2026-05-06T16:29
Anthropic is using xAI's compute moving forward I thought Elon said that Anthropic winning wasn't in the set of possible outcomes? I guess he has nothing better to do with those GPUs lol → tweet
@kunchenguid · 2026-05-05T22:36
ProgramBench is interesting - it asks models to recreate real programs (ffmpeg, SQLite, ripgrep) from scratch with no internet ... but I hope no one actually optimizes for this benchmark just think what's the solution that'll get 100% on this bench? it's a deterministic program that has the current source code of ffmepg, SQLite, ripgrep etc baked in, and just spit them out when asked → tweet
@sudoingX · 2026-05-05T19:45
anon if your kid is not growing up with a local llm on their own machine in 2026, they probably will not make it. the divide between kids who own their compute and kids who rent it is the gap of their generation. → tweet
Developer Tools & Frameworks
@steipete · 2026-05-06T06:01
Merci! imsg 0.6 + 0.7 are live 🔵 Private API bridge landed 📡 Watch/history reliability fixes 💬 Better chat + account diagnostics 🛠️ Long fallback messages decode correctly → tweet
@steipete · 2026-05-06T05:41
Me and codex were busy. 🔊 Sonos 🗃️ WhatsApp 🪶 X archive 🧰 GitHub archive 🛰️ Discord archive 🎧 Spotify 💬 iMessage 🧳 MCP to CLI 🗣️ ElevenLabs voice 🧿 second opinion Upgrading the 🦞 OpenClaw army. → tweet
@steipete · 2026-05-06T04:31
CodexBar 0.24 is live 🤖 New Windsurf, Codebuff + DeepSeek providers 👥 Copilot multi-account switching 🧹 Opt-in local storage breakdowns 🔋 Hung Codex RPC + redraw battery drain fixed → tweet
@steipete · 2026-05-06T02:33
Shipping 🛡️openclaw/fs-safe: a reusable filesystem safety primitive extracted from OpenClaw. If your Node app accepts paths from agents, plugins, uploads, configs, or users, stop treating string normalization as a filesystem boundary. Use a root handle. → tweet
@Teknium · 2026-05-06T15:41
We have greatly expanded the surface area for plugins over the last few weeks, and now you can extend LLM Inference Providers and Gateway Channels with plugins. → tweet
@Teknium · 2026-05-06T15:43
Welcome to the Hermes Agent crew! → tweet
@Teknium · 2026-05-05T22:22
RT @aijoey: In celebration of Hermes HUD Web UI crossing 1,400+ stars… v0.8.0 is live. → tweet
@badlogicgames · 2026-05-06T09:26
Using VS Code to type some code like a monkey. It's basically useless now? Auto-complete doesn't work at all or takes forever. → tweet
@badlogicgames · 2026-05-06T12:43
jfc im'a stop using agents for the rest of the week. garbage. → tweet
@sudoingX · 2026-05-05T08:40
day 6 of the month and i am already 50% through my frontier ai monthly limit. worst interaction so far was cursor. their team reached out offering credits for builders. i never had cursor installed before that dm. after the message, i installed it and emailed with my account as requested. almost 3 weeks of silence since. nobody has even seen the message. → tweet
Inference Engines & Infrastructure
@TheAhmadOsman · 2026-05-05T21:40
You don't pick an Inference Engine You pick a Hardware Strategy Then the Engine follows Inference Engines Breakdown (Cheat Sheet at the bottom) llama.cpp runs anywhere MLX Apple Silicon weapon ExLlamaV2 single RTX box go brrr vLLM default answer for prod serving SGLang vLLM but more systems-brained TensorRT-LLM maximum NVIDIA performance → tweet
@gdb · 2026-05-06T16:14
Multipath Reliable Connection (MRC): a new open networking protocol for large AI training clusters, deployed in production on our largest training clusters. → tweet
@TheAhmadOsman · 2026-05-06T06:06
There's too much alpha in cloning Sglang Mini and asking Codex Cli w/ GPT 5.5 to teach you how Inference Engines work through that cloned repo → tweet
@sudoingX · 2026-05-05T19:28
so i sat with these numbers from earlier and went looking for the turboquant speed lever everyone has been claiming. the loudest claim did not show up. gen tok/s tied with mainline at 30k loaded context, prefill -72% slower, the only real benefit was a 1.6 gb vram saving, which matters for context headroom but is not the speed boost the marketing copy promised. → tweet
@sudoingX · 2026-05-06T10:43
hot take: 90% of ai startups paying for api calls could run the same workloads locally on a single 3090 and never notice the difference. you don't need frontier pricing for tasks a 27B model handles fine. → tweet
@sudoingX · 2026-05-06T15:43
watching a 27b local model write its own benchmark report just now and i'm sitting with this for a sec. gave carnice-v2 27b (kaios SFT on qwen 3.6 dense, trained on hermes agent traces) a self-report card task, find your hardware, find your model file, find the llama.cpp commit you're running on, run a self-benchmark via curl, write a markdown report, verify it, tell me the path. it called 19 tool in 12 minutes across 42 messages, 11 terminal calls for hardware and git probing, 6 todo updates as it worked through the plan, plus one write_file for the report and one read_file to verify it, no hand holding. → tweet
Local AI & Hardware
@sudoingX · 2026-05-06T08:00
hear this. carnice v2 on hermes agent, rog scar 5090 mobile at 99% gpu, 59°C, sucking air from every direction. this is what local ai sounds like at full load. → tweet
@sudoingX · 2026-05-06T16:01
the full 12-minute run sped to 80 seconds, watch carnice v2 cycle through plan, terminal probe, todo update, write the report, then read it back to verify, until the loop closes. no edit cuts, no narration, just the model working at 5x time. the rare-bird verify moment lands in the last 20 seconds, write_file then read_file then done. closed loop on consumer hardware, you can watch it happen. → tweet
@sudoingX · 2026-05-06T07:23
if you want mac portability and you want to learn cuda, the dgx spark is the silent king nobody is talking about. 128gb unified memory in a form factor that fits on a desk corner, full cuda stack, runs nemotron 30b q8 at 56 tok/s on hermes agent, multimodal + tool calls → tweet
@sudoingX · 2026-05-06T08:16
unpopular opinion: macs hit a real ceiling for ai builders the moment you need cuda. inference on apple silicon works, mlx is real, but the moment you want kernel-level optimization, training new architectures, or agentic loops with tool calls at scale, cuda's 15year ecosystem head start becomes the wall. → tweet
@FrameworkPuter · 2026-05-06T01:37
We have 128GB Framework Desktop configurations in stock, and the inference software stack on @AIatAMD is now mature. → tweet
@kunchenguid · 2026-05-06T16:16
anyone into watching meteor showers while your agents work? one just arrived in gnhf v0.1.39 → tweet
Software Development & SaaS Displacement
@levelsio · 2026-05-05T20:36
I have 300 million users and 111,902 customers The SaaS services I replaced with AI are used by all those users and customers! → tweet
@levelsio · 2026-05-05T20:05
Extreme levels of cope in the reply section here from employees at SaaS companies that don't want to lose their job SaaS stocks went down 80% for a reason Obviously there will be SaaS that remain and new SaaS but many are just an hour or a day of work to replace with AI I already built my own screenshot service, image resizing, NSFW content moderation, community content moderation, uptime monitor with status pages, automated backup service etc → tweet
@levelsio · 2026-05-05T21:14
The advantage is mine costs $3/mo and the SaaS costs $900/mo → tweet
@levelsio · 2026-05-05T22:30
IT WORKS!! Lots of tweaking with AI but soldiers are starting to move around naturally now, really cool :O → tweet
@sudoingX · 2026-05-06T08:11
when you use these frontier models for real work every day, you would see how retarded they actually are. broken in basic ways. → tweet
@sudoingX · 2026-05-06T08:22
ill just say it. chatgpt 5.5 frontend skills are retarded. great at agentic backend, terrible at design execution. → tweet
@LinusEkenstam · 2026-05-06T13:21
Simulation. What nobody tells you. Simulation is one of the things we noticed early being a golden squeeze when working with LLMs. When building any tool that's fundamentally powered by an LLM, "what can be simulated?" is probably the first question we ask. always. → tweet
@badlogicgames · 2026-05-06T16:14
bedrock is just one clusterfuck after another. the sdk is also very very not good. i truely wonder why amazon is incapable of creating good software. → tweet
Startup & Industry News
@LinusEkenstam · 2026-05-06T15:23
The Aussie energy goes hard. → 85% YoY growth → $300bn in annual transaction volume → 200k business customers → Spike Lee directing brand films?! → $1.3bn in ARR Still most people never heard of these guys, fintech finest: → tweet
@thdxr · 2026-05-06T03:59
our teams last 7 days of spend damn gpt5.5 → tweet
@sudoingX · 2026-05-06T10:49
every ai native startup should be bringing this up in budget meetings. i say this with confidence because i have startups in my dms burning opus and sonnet credits on tasks a 27b open source model handles fine. → tweet
@levelsio · 2026-05-06T10:24
RT @marckohlbrugge: I notice similar response from taxi drivers believing human drivers will always be safer than Waymo's, engineers believ... → tweet
Security & Infrastructure
@levelsio · 2026-05-06T14:30
New fun thing I did to secure my VPS even further I installed @Cloudflare Tunnel, many of you recommended me this I already had 443 inbound firewall limited to Cloudflare's IP range, but this is even better Cloudflare Tunnel is outbound, which means it connects from your server to Cloudflare, and keeps the connection active, then if someone opens your site, Cloudflare sends you the package via the tunnel and your server responds → tweet
@levelsio · 2026-05-06T10:04
Here's exactly what I mean: On May 15, @xAI Grok will retire grok-4-1-fast-non-reasoning and API requests for them will fail I have 9 days to change the model name in 30+ sites/apps so they keep working → tweet
Community & Events
@TheAhmadOsman · 2026-05-06T18:58
Bay Area folks, I'm putting together a casual get-together in San Francisco this Saturday, May 9th Pull up to talk local AI, GPUs, infra, opensource, agents, homelabs, and whatever else the GPU-poisoned brain wants to discuss Limited capacity, first come first served → tweet
@nummanali · 2026-05-05T19:15
I spent 6 hours on feature work today Refactoring, back and forth In the end, I decided to redo it from the start I had Agent A summarise what went wrong for Agent B Second pass, 30 mins new planning, 30 mins build Much better outcome - sometimes better to let go → tweet