Executive Summary
The past 24 hours were dominated by the rapid adoption and evaluation of AI agent frameworks, with xAI's Grok Build quietly opening to X Premium+ users and Hermes Agent gaining significant traction for local and cloud deployments. Developer sentiment increasingly favors Cursor over Claude Code for agentic coding, citing faster agent loops and tighter context handling. On the hardware front, users demonstrated that 10-year-old GTX 1080 cards can run current open-weight agentic models with 600K+ token context, challenging the narrative that local AI requires expensive hardware. Open-source AI advocacy intensified, while a new MIT-licensed talking-avatar model from LongCat emerged as a potential SOTA release.
Key Events
- xAI's Grok Build appears accessible to X Premium+ users without SuperGrok subscription — users report the CLI agent works out-of-the-box, possibly due to loose gating or unannounced tier promotion. → link
- Hermes Agent sees mass adoption wave — EU Bitwarden integration, team deployments with shared memory, and confirmed functionality on DGX Spark with Qwen models overnight. → link
- Cursor outperforms Claude Code in head-to-head agentic coding tests — developers report Anthropic's tool produces half-complete, untested output while Cursor's agent loop is faster and more rigorous. → link
- LongCat releases open-source talking-avatar model (MIT-licensed) — potential SOTA, immediately deployed on Hugging Face Spaces for AI tutors, dubbing, and talking-head agents. → link
- GTX 1080 (10-year-old card) runs agentic models with 656K context — Gemma 4 e4b at q4 quant achieves 42 tok/s on 8GB VRAM, proving local AI accessibility on obsolete hardware. → link
- Hugging Face shares hardware survey results from 300,000 AI builders — data on community GPU/compute configurations released. → link
- AI cost crisis hits tech giants as employee "tokenmaxxing" backfires — Tom's Hardware reports on ballooning inference costs. → link
- Composer 2.5 achieves more post-training RL compute than base model pretraining — signaling a shift in how model capability is scaled. → link
Analysis
Patterns: - Local AI credibility surge: Multiple independent voices (sudoingX, TheAhmadOsman) are pushing back against the narrative that local AI requires expensive hardware, with empirical benchmarks as proof. The GTX 1080 thread is particularly significant — it provides a replicable recipe (Q4 quant + Q4_0 KV cache + arch-specific llama.cpp build) that undermines GPU vendor upgrade cycles. - Agent framework convergence and divergence: Hermes Agent is consolidating as the preferred harness for both local and cloud workflows, while Grok Build enters as a new contender. Meanwhile, the Cursor vs. Claude Code comparison reveals that model quality alone doesn't determine agent utility — the harness, loop speed, and context management matter as much or more. - Open-source advocacy sharpening: TheAhmadOsman's "Open-source AI must win" framing and multiple references to open-weight model accessibility suggest the community is coalescing around an existential narrative opposing closed-source dominance by OpenAI/Anthropic.
Escalation: - Grok Build's apparent loose gating to X Premium+ users could accelerate xAI's developer footprint rapidly if confirmed. - Cursor's performance edge over Claude Code, if sustained, threatens Anthropic's subscription tooling moat.
What to watch next: - Whether xAI confirms or closes the Grok Build access gap for Premium+ users. - Agentic coding benchmarks on older hardware (sudoingX promised multi-file refactor tests next). - Hermes Agent's continued enterprise/team adoption and whether shared memory features scale. - LongCat talking-avatar integrationinto coding agents and commercial products.
Tweet Feed
AI Agent Frameworks & Tools
@sudoingX · 2026-05-24T17:26
first impressions of grok build by xai and i already love the vibe. clean tui that's fast and interactive clickable, the token allocation diamond visual is genuinely a nice touch. signed in with x premium+ and it just worked out of the box. 512k context window, 28 tools, auto compact at 85%. ill keep testing and reporting back. → tweet
@sudoingX · 2026-05-24T16:44
i was about to subscribe to supergrok this week just to try grok build. the public docs say supergrok heavy is required. instead i signed into the grok cli with my x premium+ account and it just works. my config shows installer = "internal" and the cli is pointed at cli-chat-proxy. an agent that read my config thought this meant deliberate partner tier access but i never received any partner link, never got an early access email, never asked anyone for access. just installed the cli, signed in with x, got grok build. either the public docs are outdated, the gate is loose, or x premium+ accounts are getting promoted access without it being announced anywhere. anyone else on x premium+ should install the grok cli and report back if grok build works without supergrok. let's see if it's just my account or the gate is open. → tweet
@sudoingX · 2026-05-24T16:59
anyone at xai on the grok build team, ping me for my peace of head, peacefully testing the model and grok keeps confidently telling me this same thing every session. that someone on your side provisioned my account with internal/partner tier (tier 4 + team + grok-cli scope) tied to nousresearch contributor status + the active grok provider pr, parallel internal distribution path that lets normal x login unlock the full agent binary without going through the supergrok heavy route. either xai actually did this and grok is leaking the lore, or grok is hallucinating a very specific story about me. both are interesting, just tell me which. ill keep building either way. → tweet
@sudoingX · 2026-05-24T16:08
look at this. opus 4.7 max thinking on claude code, the moment my cursor reviewer pushes back. realizes it's fucked. same opus 4.7. same max thinking. same task. caught by the same model running in cursor's harness. cursor wins every single time and it's not close. claude code feels like a q4 quant of the same opus 4.7 max these days. half the rigor, missed context, sloppy pre-audits. idk if anthropic is nerfing the subscription model or if the harness is being built by retards. cursor has nailed something the claude code team hasn't, and i'm having to rework everything now. cursor is my primary frontier now. it's what i always wanted claude code to be. → tweet
@sudoingX · 2026-05-24T17:37
this is so annoying. most of the work done by claude code, my cursor reviewer finds it half complete and untested every single time. absolute slap on anthropic engineers face by cursor. → tweet
@sudoingX · 2026-05-24T13:56
10 days since cursor granted the 10k credits. $1,259 burned so far. most of it landed in the last 48 hours. eight days of light use, mostly the fallback when my other subs hit rate limit walls. then the last two days cursor went from fallback to daily driver, and i burned more in 48 hours than the previous 8 days combined. the underlying foundation is built different. claude code and codex are both good but cursor's agent loop is faster, the context handling is tighter, the iteration cycle has less friction. speed specifically. cursor moves at a pace the other tools just don't match, and once you feel it you stop wanting to go back. 87% of the credit pool still in reserve. acceleration window wide open. → tweet
@Teknium · 2026-05-24T10:57
EU Bitwarden users now have access to it in Hermes Agent, and self hosting as well! → tweet
@Teknium · 2026-05-23T19:46
.@NetworkChuck's overview of what makes Hermes Agent unique is really well done. He's clearly been a power user and done his homework in this video. Highly recommended if trying to decide if you should try Hermes Agent! → tweet
@Teknium · 2026-05-24T04:12
Welcome to the Hermes Agent crew! → tweet
@Teknium · 2026-05-24T11:51
RT @cptn3mox: Moved over to hermes for a day and man the way it writes replies is so clean? They're concise yet elaborate, bolds are right… → tweet
@Teknium · 2026-05-24T14:48
RT @libapi_: 这两天给 Hermes Web UI 做了几波小更新 v0.5.34 主要是把历史记录、模型配置、Kanban 和上下文统计弄干净。v0.5.35 面板更新： 装了个性能监控ui… → tweet
@gdb · 2026-05-24T02:56
under appreciated that codex is open source → tweet
@gdb · 2026-05-24T17:18
self improvement prompt for codex → tweet
@thdxr · 2026-05-24T17:38
effect is actually bad for agents → tweet
Local AI & Hardware Benchmarks
@sudoingX · 2026-05-23T20:14
gtx 1080 8gb of vram launched may 2016. card turns ten this month. just ran three current open weight agentic models on one and the smallest of them fit 656,000 tokens of context at 38 tok/s gen speed. on a pascal arch card with no tensor cores. on 8gb of gddr5x that the discourse keeps telling me is unusable. three models, same hardware, same locked flags. qwen3 8b, qwen 3.5 9b, gemma 4 e4b. q4_k_m quant across the board. q4_0 kv cache, flash attention on, llama.cpp built for sm_61. one line setup. results — vram ceiling: qwen3 8b 78k, qwen 3.5 9b 248k, gemma 4 e4b 656k. gen tok/s at small context: 31.71, 29.91, 42.13. gen tok/s at ceiling: 31.78, 29.62, 38.73. gemma sweeps every category. 2.6x more context than qwen 3.5 9b, 8.4x more than qwen3 8b, 30% faster at the ceiling. → tweet
@sudoingX · 2026-05-24T11:51
so yesterday i dropped the bench numbers and what fits. today is the actual agent running on this 10 year old gpu card. qwen3 8b q4_k_m on a gtx 1080 8gb. hermes agent loaded with full tool set, browser controls live, nvtop pinned at 100% gpu 7.5gb of 8gb vram occupied. the unsloth weights pulled directly from huggingface, q4 quant, llama.cpp built for sm_61 (the pascal compute capability that everyone forgot exists). 31 tok/s gen speed, faster than most people read. this is what happens after the bench. raw perf was the receipt for what fits. now we test what actually works. agent loops, tool calls, real coding tasks coming next. → tweet
@sudoingX · 2026-05-24T14:22
running ai locally is not as hard as people put it out to be. i'm running current open weight agents on a 10 year old gtx 1080 pulled from a drawer. one gguf, one llama.cpp build, one agent harness. that's the whole stack. cloud has its lane. the floor for local is way lower than the gatekeepers pretend. own your cognition. start with the hardware you already have. the rest is iteration. → tweet
@sudoingX · 2026-05-23T21:14
check your drawers first. the replies are already turning into a hardware roll call of old silicon people forgot they own. half of what gets called obsolete is under optimized, the rest is people who never tried. q4 quants, q4_0 kv cache, llama.cpp built for your arch. that's the whole ritual. open the drawer before you open the wallet. → tweet
@sudoingX · 2026-05-24T10:02
i will keep saying it. dgx spark with hermes agent /goal is the most underemployed combo in local ai right now. just switched to qwen 3.6 35B-A3B at Q8 with MTP. 262K context. the agent ran overnight autonomously and is now self updating its memories from last night's session. i woke up and it's already caught up on what changed. hermes agent just works on local models. no hacks. no workarounds. you set a /goal, you sleep, you wake up to progress. the memory system means it picks up where it left off. this is what agentic local ai is supposed to feel like. → tweet
@TheAhmadOsman · 2026-05-24T18:35
Don't know where to start with Local AI? Read my Local LLMs From Zero to Hero series. It covers: Hardware, Software, Models Mechanics, Everything else necessary. Needs no prior experience. Easy to understand for any background. Local / Opensource AI FTW → tweet
@TheAhmadOsman · 2026-05-24T17:23
You should buy an RTX 3090 and learn how to run models locally. The elite don't want you to know this but running local models is hella easy, performant, and cheap nowadays → tweet
@sudoingX · 2026-05-23T20:31
when you somehow scale more compute on post training rl than the base model spent on pretraining. that's it. composer 2.5 happens. → tweet
Open Source AI & Models
@victormustar · 2026-05-24T10:16
New: LongCat just dropped an excellent open-source talking-avatar model (probably SOTA) + MIT licensed 🔥 Made a Hugging Face Space for it and it's very impressive. So many cool products to build with it: AI tutors with a face, dubbing pipelines, talking-head coding agents (imagine Claude Code with a face), NPC dialogue, etc... → tweet
@TheAhmadOsman · 2026-05-23T21:41
Opensource AI should remain usable, understandable, reproducible, locally deployable, economically viable, and community-governed even if today's dominant labs, foreign labs, hardware vendors, cloud platforms, or open-weight model providers change direction or disappear. → tweet
@TheAhmadOsman · 2026-05-23T20:11
Opensource AI MUST WIN. OpenAI / Anthropic winning is - At best a world we can tolerate - At worst a dystopia. We lose in all scenarios, the degree of that loss is a matter of how things play out + being UNDER THEIR MERCY. This is our MOST IMPORTANT FIGHT. Existential. → tweet
@badlogicgames · 2026-05-24T17:44
recommended reading. i wanted to write this blog post, describing how we OSS in pi, for literal months. now i don't have to anymore. thanks, creator of pi. → tweet
@badlogicgames · 2026-05-24T17:23
RT @mitsuhiko: Has been a while since I wrote about agentic engineering, so this time around some learnings of maintaining Pi as a junior m… → tweet
@badlogicgames · 2026-05-24T17:21
RT @ClementDelangue: 300,000 AI builders filled their hardware profile on @huggingface and we're sharing the results: → tweet
Developer Tools & Engineering
@steipete · 2026-05-24T02:54
I always wanted a GitHub dashboard: See my repos, open Issues/PRs, what version I released last, how many commits since last release. So I built one for everyone. → tweet
@steipete · 2026-05-23T22:03
I'm refactoring an older part of the codebase (subagents) that touches a lot of code, and autoreview is running for 5h already and fixing tons of issues. → tweet
@steipete · 2026-05-23T23:40
codex... made a smiley? :) → tweet
@hnasr · 2026-05-24T14:01
Root Cause: Stories and Lessons from Two Decades of Backend Engineering Bugs. No best practices. no recommendations. just stories of the most interesting bugs and how I solved them. → tweet
@badlogicgames · 2026-05-24T01:05
RT @cnakazawa: Cloudsail: Instant Sandboxes for Coding Agents. Create a new Cloudflare Sandbox for each task with a shell, Codex and GitHub… → tweet
@jezell · 2026-05-24T06:58
RT @shreemanarjun: 🚀 Nitro v0.4.3 is live! Fixes JVM crashes, upgrades CI (--no-ui), & adds a new clean command. 🔥 Plus, work has officially… → tweet
@jezell · 2026-05-24T00:38
RT @seconds_0: So, OpenAI batch API is 50% off and says it can be up to 24 hours to respond. Want to know my practical average over 2800 (l… → tweet
@jezell · 2026-05-24T00:50
RT @JoeBeOne: Killing long-lived developer API and GPG keys was a massive victory for the architecture of the Internet. But attackers didn'… → tweet
@RydMike · 2026-05-23T19:29
And made with Flutter, one of my favorite #FlutterDev apps! 💙 → tweet
@Teknium · 2026-05-24T18:59
RT @DavidOndrej1: > learn how to use tmux > trust me → tweet
@sudoingX · 2026-05-24T14:08
settle it. your AI usage split in may 2026. → tweet
AI Industry & Trends
@jezell · 2026-05-24T14:08
RT @tomshardware: AI cost crisis hits tech giants as employee 'tokenmaxxing' backfires → tweet
@levelsio · 2026-05-24T11:48
I know Anthropic has a GPU shortage but every day forcefully putting my effort back to medium feels...well....annoying → tweet
@kunchenguid · 2026-05-24T04:02
i've been in both management and IC position at senior levels so I have a lot to say about this. [...] especially now, ICs have a lot more time and freedom to play with new technology, solve interesting problems and gain a massive leverage by utilizing AI effectively - this is extremely important skill for staying relevant in the coming era. in the AI era, ICs are the ones getting the biggest boost in leverage. in the past couple of years, ICs have gone from having small assistance from copilot code complete, to getting some tasks done by agents, to now running 10-100 agents at the same time. middle managers are roughly doing the same things as 10 years ago. → tweet
@thdxr · 2026-05-24T01:41
any lack of polish now gets equated to ai slop. all fields: software bugs, bad movie, bad video game. before the audience would make a judgement about your skills, which was tolerable because you can get better. now they make a judgement about your character (lazy/fraud/etc) → tweet
@TheAhmadOsman · 2026-05-24T00:06
Karpathy's title at Anthropic is Member of Technical Stuff. Things you get on with so you can get your hands on some Compute nowadays I guess. Wild ngl → tweet
@levelsio · 2026-05-24T09:14
That's a really outdated mindset. Japan doesn't even have its own LLM like DeepSeek. Its biggest model is Rakuten AI which is a finetuned version of Chinese DeepSeek. The cope about China not innovating can't last forever → tweet
@levelsio · 2026-05-24T09:09
Just a decade ago Japan would have invented this, they love cats and tech. But this is a Chinese device by a Chinese startup running a Chinese AI model. The center of innovation in Asia has shifted to China, not just production! → tweet
@badlogicgames · 2026-05-24T14:12
RT @dillon_mulroy: writing code with ai and deeply understanding what you're building and how it works are not (and should not be) mutually exclusive → tweet
Robotics & Hardware Hacking
@badlogicgames · 2026-05-24T11:13
the robot's brain is done. pi + elevenlabs plus phone sensors/cameras + memory system. will build a cardboard chassis for phone/electronics tonight, put it on top the robot legs, then the MVP is done. love living in the future. → tweet
@badlogicgames · 2026-05-24T18:11
SHE'S A GENIUS! now i just need to mount the phone with some cardboard, solder a shorter usb cable and the MVP is done. best of all: i can take this as a base design into fusion! very easy to work off of an easily measurable cylinder! kisses on the forhead @wheresmyfu_pony → tweet
@badlogicgames · 2026-05-24T08:46
robot intensifies → tweet
@badlogicgames · 2026-05-24T00:58
RT @badlogicgames: proof of concept done! - remove soic from original robot - wire up h-bridge with ft232h - connect ft232h to android pho… → tweet
@badlogicgames · 2026-05-23T23:57
RT @badlogicgames: it's alive, driven via my mac through a little C programm pi wrote. now i just need to port this to Android and i'm good… → tweet
@badlogicgames · 2026-05-24T00:03
RT @badlogicgames: witness the true power of agents. gpt (via https://t.co/oUoqqL9hAp) just wrote me a web usb controller instead of the cl… → tweet
@badlogicgames · 2026-05-24T00:58
RT @antirez: I finally found the solution I wanted to the old/new editing problem. And it is a solution that at the same time works extre… → tweet
UI/UX & Creative AI
@LinusEkenstam · 2026-05-24T18:52
Unreal 6 wild reaction in Paris, volume up, it's absolutely nuts → tweet
@MengTo · 2026-05-24T03:53
A single prompt to turn images into videos for your landing pages → tweet
@badlogicgames · 2026-05-24T17:22
looks at safari (iOS) i agree tho, native's been over for a long time. i hope i never have to touch win32/appkit/uikit/compose/whatever ever again. except maybe qt. that was good. sorta. → tweet