Executive Summary
The past 24 hours saw significant momentum in agentic AI tooling: OpenAI's Codex expanded to mobile and multi-device workflows, while the Hermes Agent ecosystem grew with xAI Grok integration and strong community adoption. Local LLM inference benchmarks for AMD hardware emerged concretely—an RX 7900 XTX hitting 68.79 tok/s on Qwen 3.6 35B-A3B via Vulkan—filling a long-standing gap in AMD performance data. A widely-shared thread from @sudoingX crystallized best-practice infrastructure for agentic systems (Tailscale, tmux, Git-as-memory-layer), reflecting the community's shift from model obsession to foundational DevOps-style agent infrastructure. On the model front, EAGLE-3 speculative decoding drafters are being trained for Qwen-3.6-27B, and a pragmatic "top 3 models worth building on" ranking provided rare honest benchmark signal.
Key Events
- OpenAI Codex expands to mobile and multi-device workflows — @gdb (OpenAI) demonstrated building from a phone via ChatGPT app and linking devices together with Codex, signaling a push toward ambient, anywhere development. → link
- Hermes Agent integrates xAI Grok subscriptions and X Search — Teknium announced Hermes Agent now works with Grok via OAuth, giving agents X Search, image/video gen, and voice capabilities. → link
- AMD local LLM benchmarks go public — An RX 7900 XTX running Qwen 3.6 35B-A3B at iQ4 achieved 68.79 tok/s on Vulkan, the first concrete benchmark in an open call for AMD inference numbers. → link
- "Five foundations" for agentic systems goes viral — @sudoingX detailed Tailscale, Termius, tmux, private Git repos, and scripting as non-negotiable infrastructure for agent workflows, sparking extensive community discussion. → link
- EAGLE-3 speculative decoding drafter in development for Qwen-3.6-27B — @Ex0byt is training/distributing an EAGLE-3 drafter to accelerate Qwen inference, highlighting specdec as an active optimization frontier. → link
- Top 3 local models worth building on ranked — Step-3.5 Flash 121B MoE, Qwen 3.6 27B Dense (Q4), and Nemotron 3 Nano Omni 30B-A3B named as the only models worth hardware investment across 3090→DGX Spark range. → link
- Codiff 0.1 released for local code reviews with optional LLM walkthroughs — Positioned as companion tool for developer workflows alongside Codex. → link
- Node.js minimum version bumped to 22.19.0 — @badlogicgames announced the change driven by undici-related compatibility, affecting libgdx/processing users. → link
- llama.cpp from source advocated over LM Studio/Ollama for serious local inference — @sudoingX argued that compiling from source gives granular control over CUDA arch, quantization kernels, flash attention, and context size with no abstraction walls. → link
Analysis
Patterns: The conversation has decisively shifted from "which model is best" to "how do I actually run agents reliably." The five-foundations thread, the Git-as-memory-layer incident (nearly losing a 97MB agent session), and the AMD benchmark call all point to infrastructure maturity overtaking model hype as the primary concern for practitioners.
Escalation: Local inference on non-NVIDIA hardware is gaining measurable traction. The Vulkan-based AMD benchmark is a concrete data point in what has been an evidence desert. Meanwhile, the agent ecosystem is bifurcating between OpenAI's Codex (cloud-integrated, mobile-first) and open frameworks like Hermes (local, skill-based, multi-model). Both are accelerating.
De-escalation: The "Bun to Rust" and Jane Street/OCaml discussions suggest the language wars are cooling into pragmatic pluralism rather than tribalism.
What to watch next: - AMD Strix Halo (128GB) benchmarks from @sudoingX when hardware arrives—could reshape local inference cost assumptions. - EAGLE-3 drafter release for Qwen-3.6-27B—speculative decoding could materially change throughput economics. - Codex mobile workflow adoption rates—whether phone-based development becomes a real pattern or a demo novelty. - Whether AMD community benchmarks aggregate into a usable reference table.
Tweet Feed
AI Models & Benchmarks
@sudoingX · 2026-05-17T05:10
i've run a stack of models across a single 3090, a 5090, and a 128GB DGX Spark. exactly three are worth building on. the honest list. the three worth it: > 1. StepFun Step-3.5 Flash, the REAP pruned 121B MoE (Q6, DGX Spark)... > 2. Qwen 3.6 27B Dense, Q4 (single RTX 3090)... > 3. NVIDIA Nemotron 3 Nano Omni, 30B-A3B (DGX Spark)... → tweet link
@Ex0byt · 2026-05-17T17:19
the different flavors of specdec, and why I'm trying produce a Qwen-3.6-27b EAGLE-3 drafter for ya'll → tweet link
@Ex0byt · 2026-05-17T16:07
distilling/training/testing an EAGLE-3 drafter for Qwen-3.6-27B 🦥. A Dell Pro Max w/ GB300 would fix me rn… help😇 → tweet link
@sudoingX · 2026-05-17T08:07
first fresh AMD number in and it's a good one. an RX 7900 XTX, 24GB, running Qwen 3.6 35B-A3B at iQ4, 68.79 tok/s generation on vulkan. → tweet link
@sudoingX · 2026-05-17T06:46
if you are running local models on AMD right now, R9700, Strix Halo, a 7900 XTX, any RDNA card, on a ROCm or Vulkan build, drop your numbers in the replies. model, quant, the card, tok/s. → tweet link
@sudoingX · 2026-05-17T06:35
honest answer: not yet. all of it ran on nvidia... i don't have an AMD card in hand... AMD is the next gap to close. i've got a 128GB Strix Halo box inbound, and the day it lands, ROCm and Vulkan builds get the same honest benchmark pass. → tweet link
@badlogicgames · 2026-05-17T15:57
labs will do anything to smooth out that jagged intelligence. → tweet link
@badlogicgames · 2026-05-17T16:02 (RT @antirez)
Fixed two subtle inference errors from 2bit Flash (still not pushed), and removed the broken tests after checking... → tweet link
@TheAhmadOsman · 2026-05-17T16:24 (RT @ChinmayKak)
very interesting and offbeat paper from hunyan llm team where they introduce a third reference model in the standard OPD ob… → tweet link
@louszbd · 2026-05-17T12:10 (RT @ericavaneee)
We built TERMS-Bench, a three-tier benchmark for LLM agents in real-world economic negotiation. No LLM-as-judge, no outcome… → tweet link
@TrungTPhan · 2026-05-17T17:49 (RT @bearlyai)
Cerebras CEO Andrew Feldman talks about why there are so few young semiconductor chip founders: "The silicon industry is not… → tweet link
Agentic AI Infrastructure
@sudoingX · 2026-05-17T09:54
anyone thinking about, learning, or already working with agentic systems, you should know this. the first few steps of your setup matter more than any model or framework you pick later... the foundation nobody posts about: 1. tailscale... 2. termius... 3. tmux... 4. a private git repo... 5. script everything from day one. → tweet link
@sudoingX · 2026-05-17T16:27
someone asked me to elaborate on #1, tailscale... an agent that works across machines has to reach those machines. without a tailnet you are fighting public IPs, port forwarding, firewall rules, NAT, jump hosts... → tweet link
@sudoingX · 2026-05-17T16:21
and here is proof the five hold. a devops engineer in the replies runs every one of them, then takes them up a tier, hermes agents in a GKE cluster with self hosted models behind them. that is the part most people miss. the foundations do not change when you scale. → tweet link
@sudoingX · 2026-05-17T15:54
i nearly lost a 97MB agent session. it grew so large it will not reload... the code was committed to a git repo, so it is still in there and recoverable... what i am scrambling to recover is the context i left in a chat window. foundation 4, not as advice, as something that happened to me hours ago. the repo is the memory layer. the chat is not. → tweet link
@sudoingX · 2026-05-17T16:06
the replies are doing exactly what i hoped, adding tools i didn't list. this one is worth catching: reptyr. start a long process bare, outside tmux, then realize you need it inside a session, reptyr reparents the running process into tmux after the fact. → tweet link
@sudoingX · 2026-05-17T03:44
for months my ceiling was a few commits a day. not because the ideas ran out, because the frontier agent budget did... the spike at the end is the last few days, since the cursor $10k credits landed and the budget stopped being the wall. → tweet link
Hermes Agent Ecosystem
@Teknium · 2026-05-16T20:37
You can now use SuperGrok direct subscriptions, and also X Premium+ Subscriptions to access Grok, X Search, Image and Video Gen, and Voice! Check out the main post for info on the new X Search tool that your agent will automatically have if using Grok oAuth login! → tweet link
@Teknium · 2026-05-17T18:21 (RT @NiteshTechAI)
LA Hermes Agent meetup is happening. Live demos. Real automations. Builders who actually ship. → tweet link
@Teknium · 2026-05-17T11:06
Wen gadot skill? → tweet link
@Teknium · 2026-05-17T05:51 (RT @Saboo_Shubham_)
This is bigger than YOU think. Hermes Agent now works with xAI Grok Subscription. I just added a new X Research Agent… → tweet link
@Teknium · 2026-05-17T11:12 (RT @tonysimons_)
🧠 Hermes Agent Tip of the Day:
hermes send— Your shell scripts can now text you. 🚫 No agent. 🚫 No LLM. 🚫 No gatewa… → tweet link
@Teknium · 2026-05-17T02:23 (RT @Sentdex)
162M tokens into MiniMax M2.7 running locally w/ Hermes for work, data analysis, and general daily Q&A. I really cannot get ov… → tweet link
@Teknium · 2026-05-17T02:24
So much good content around Hermes! → tweet link
@Teknium · 2026-05-16T23:58
@Teknium · 2026-05-17T18:13
Wonder if he tried Hermes Agent → tweet link
Developer Tools (Codex, Codiff, llama.cpp)
@gdb · 2026-05-17T05:24
you can just build things from your phone, with Codex in the ChatGPT app → tweet link
@gdb · 2026-05-17T16:19
link together your devices with Codex to develop from anywhere, anytime → tweet link
@gdb · 2026-05-16T23:25
tokens are rapidly becoming the universal input for solving problems → tweet link
@gdb · 2026-05-16T20:34
keep the feedback coming, team will keep shipping → tweet link
@steipete · 2026-05-16T20:27
deslop your Claude code if you haven't yet switched to Codex. → tweet link
@steipete · 2026-05-17T03:43 (RT @fitblake2)
codex fixed my claude code set up and now its 100x better. Maybe i just need codex. → tweet link
@steipete · 2026-05-17T06:03 (RT @cnakazawa)
Codiff 0.1 is out — Fast Local Code Reviews, Optional LLM Walkthroughs, Inline Review Comments. This is the best companio… → tweet link
@steipete · 2026-05-17T10:30 (RT @joshavant)
Just using Crabbox and it needed me to perform a browser-based OAuth flow.... so my clanker opened up a browser window w/ an… → tweet link
@sudoingX · 2026-05-17T06:23
people keep asking what engine i use. no lm studio. no ollama. i compile llama.cpp from source every time for personal inference... if you're serious about local inference, start at source level. when you compile from source you control everything. → tweet link
Software Development & Languages
@badlogicgames · 2026-05-17T18:39
People of [libgdx]. Due to recent Node changes related to undici, we need to set the minimal Node version to 22.19.0 from 20. I'm sorry. Welcome to the future. → tweet link
@badlogicgames · 2026-05-17T16:05 (RT @davidcrawshaw)
While the industry is pouring resources into programs without GC (rust), I think the Jane Street OCaml folks have it fig… → tweet link
@badlogicgames · 2026-05-17T16:04 (RT @rovarma)
Me: we're running into an issue on Linux with dbus, we think it's related to
github issue. Claude: <long, detailed, pla… → tweet link
@thdxr · 2026-05-16T22:57
the bun to rust thing i've had little feeling around it besides "huh that's interesting wonder how it'll go" and i'm 100x more impacted than the avg person. but so many people here so worked up over it on both sides of the argument i don't get it → tweet link
@badlogicgames · 2026-05-17T09:38 (RT @ryoppippi)
ccusage、過去最大級のupdateをしました (ccusage major update released) → tweet link
@hnasr · 2026-05-17T14:01
You cannot make a good implementation until you first make a naïve one. Of course, you can do what others told you is a good implementation but you will be blind and confused. That is why listening to best practices is moot until you get the scoop of what the bad practices were. → tweet link
AI Applications & Commentary
@TrungTPhan · 2026-05-17T01:34
Singapore's Foreign Minister built a NanoClaw AI agent on a Raspberry Pi so he could "learn by doing" (he chats with it on WhatsApp for meeting prep)... he says AI agents can get expensive and will be more so when labs stop subsidizing tokens. → tweet link
@TrungTPhan · 2026-05-17T15:53
"He's about to get an entry level white-collar job, release the Claude plugin that automates and eliminates the job listing." → tweet link
@levelsio · 2026-05-17T11:10
What in the AI noise reduction is happening in this video? It's unlistenable to me. If you remove all background noise it just sounds like fake dubbed AI voices, so bad → tweet link
@TrungTPhan · 2026-05-17T17:54 (RT @pnegahdar)
HN loved this one: Nano, a 200 line agent harness with zero dependencies. Supports: Skills, Claude/Agents.md, sessions, re… → tweet link
Hardware & DIY Electronics
@TheAhmadOsman · 2026-05-17T00:43
Buy a GPU predicted this a year out btw. Anyone remembers the $6,000 RTX PRO 6000 I used to share on the timeline? Good times → tweet link
@badlogicgames · 2026-05-17T18:41
today i rescued 1 Tonibox and 1 bicycle LED with another bit of soldering. superficial electronics knowledge is a super power. learn to solder! → tweet link