← Tech / AI / IT Monitor Index Tech / AI Generated 2026-06-24 19:30 UTC

Tech / AI / IT Monitor

June 24, 2026 · Based on tweets from the last 24 hours · 124 tweets analyzed · model: ollama-cloud/glm-5.1:cloud

Executive Summary

The past 24 hours saw major AI product and infrastructure announcements: OpenAI's @gdb unveiled GPT-5.5 Instant improvements and a new custom chip codenamed Jalapeño purpose-built for LLM inference. Nous Research's Hermes Agent dominated attention with a new /learn skill-acquisition feature, animated agent pets, and a resounding leaderboard lead amid a public spat with OpenClaw over bloat and security. Figma Config 2026 kicked off with the launch of Figma Motion, enabling in-app animation with code/video export. Meanwhile, GLM-5.2 emerged as a surprisingly strong contender (ranked above Opus 4.8 by some), and Andrej Karpathy articulated a compelling vision for the third major LLM UI paradigm — persistent, asynchronous, org-wide AI entities.

Key Events

Analysis

Agent frameworks are the new battleground. The Hermes vs. OpenClaw confrontation reveals a maturing ecosystem where differentiation now centers on security posture (deny-by-default vs. kitchen-sink packages), local inference compatibility (tool-call repair), and cost efficiency (token throughput). The market is fragmenting between "batteries-included" and "minimal-privilege" philosophies, with both camps claiming dominance via token-volume metrics.

The UI paradigm shift is real and accelerating. Karpathy's articulation of the third LLM paradigm — persistent org-wide agents — aligns with Hugging Face's Moon Bot and Hermes's cron/gateway system. The industry is converging on agents that don't just chat but operate continuously with scheduled tasks, tool integrations, and shared organizational memory.

Open-source and local inference are gaining conviction. Multiple voices (@TheAhmadOsman, @sudoingX, community demands for Qwen 3.7) signal rising confidence that local-first models can handle agentic workloads — provided the harness layer is competent. The Jalapeño chip announcement suggests even large providers are investing in inference-specific hardware, which could further compress local/cloud cost gaps.

What to watch next: (1) Whether Hermes's /learn feature sustains its quality at scale and whether community-built skills proliferate; (2) Figma Motion's impact on design-to-code workflows post-Config; (3) GLM-5.2's full benchmark suite once widely available; (4) OpenAI's Jalapeño chip specs and deployment timeline; (5) Whether Google's talent-retention crisis deepens following the Workspace CLI firing and Addy Osmani's departure.

Tweet Feed

AI Models & Benchmarks

@gdb · 2026-06-24T18:09

Big improvements to GPT-5.5 Instant, including being much more fun to talk to. Give it a try: → tweet link

@gdb · 2026-06-24T15:46

Introducing Jalapeño — designed from scratch for LLM inference over nine months, accelerated by our models. Perf per watt looking incredible. → tweet link

@TheAhmadOsman · 2026-06-23T23:18

GPT 5.5 > GLM 5.2 / But / GLM 5.2 > Opus 4.8 → tweet link

@Ex0byt · 2026-06-23T22:06

Update: the road to GLM-5.2: we're getting there, folks! non-quantized, non-pruned DeepSeek-v4-Flash. 11tok/s on a single DGX Spark. sglang inference + custom mega-kernel. Pure beauty. → tweet link

@louszbd · 2026-06-23T21:05

Made it to SF! The love for GLM-5.2 has been incredible. We are bringing team out for the AI Engineer World's Fair, where we'll be sharing some of our recent work. It's our first time showing up in the Valley. → tweet link

@cooltechtipz · 2026-06-24T07:55

GLM-5.2 vs Claude Opus → tweet link

@cooltechtipz · 2026-06-24T11:54

AI progress → tweet link

@tinygrad · 2026-06-23T19:08

Today, if you were writing a bunch of kernels, what would you reach for? Raw CUDA? tile-lang? Triton? ThunderKittens? → tweet link

Hermes Agent & Agent Framework Wars

@Teknium · 2026-06-23T21:07

Hermes can now LEARN from any source or set of sources, build a skill, test it live, and crystallize new learnings. Just run /learn and pass it sources, past sessions, URLs, docs, whatever you think will help it learn, and it'll go from 0 to 1 to create you a skill! → tweet link

@Teknium · 2026-06-24T00:39

Not everything needs an llm involved — Hermes can run regular scripts through its cron system and use its gateway to update you on outcomes without the agent burning money 😇 → tweet link

@Teknium · 2026-06-24T03:07 (RT @NousResearch)

Your Hermes Agent can now adopt an animated pet: a small sprite that reacts to what the agent is doing (idle, running a task…) → tweet link

@sudoingX · 2026-06-23T19:25

this isn't the voice of an open-source contributor. this is the openai paycheck talking… hermes agent ships 225 [packages], every one pinned to an exact version… 1.03 trillion tokens in a single day. more than the entire rest of the top five combined. → tweet link

@sudoingX · 2026-06-24T13:10

read this if you think your local model is too dumb for agentic work. your bloated harness might be eating its tool calls for breakfast… most bloated harnesses just trust the inference server to parse tool calls. local servers like llama.cpp and vllm hand back malformed ones all the time. → tweet link

@sudoingX · 2026-06-24T07:27

if you're still running openclaw bloat for agentic work in june 2026, drop it today. you don't need the bloat. you need best #1 harness. hermes agent. stop making this harder than it is. → tweet link

@sudoingX · 2026-06-24T14:59

i live and breathe agents. every conversation, every post, every decision i make runs through them… the thinking is 100% mine. that's not a confession, it's the future you're too scared to touch. → tweet link

@sudoingX · 2026-06-24T12:43 (RT)

somewhere in alibaba there is an engineer sitting on the exact model i need. qwen team, a 3.7 27b checkpoint. → tweet link

LLM Paradigms, Research & Technique

@karpathy · 2026-06-23T22:26

This is the 3rd major redesign of LLM UIUX. The first paradigm was that the LLM is a website you go to, the second was that it is an app you download. This third one is that it is a self-contained, persistent, asynchronous entity with org-wide tools and context, working alongside teams of humans. → tweet link

@badlogicgames · 2026-06-23T22:24

recommended reading. deepmind's new AI control roadmap. looks like they've given up on solving the lethal trifecta directly. the new direction seems to be a tower of LLMs. → tweet link

@cooltechtipz · 2026-06-24T15:40

How KV cache optimization makes LLMs respond faster. → tweet link

@cooltechtipz · 2026-06-24T09:44

How multi-agent systems divide and coordinate work. → tweet link

@cooltechtipz · 2026-06-24T05:27

Guide to long-context AI. → tweet link

@jsuarez · 2026-06-24T15:13 (RT @daphne_cor)

We've seen that self-play is an effective training strategy for driving policies from vectorized BEV features. → tweet link

@juliarturc · 2026-06-24T17:16

Refreshing when smart people like @neetcode1 call out the emperor for being naked. The antidote to hype is specificity, just ask "But how EXACTLY?". → tweet link

@thdxr · 2026-06-24T02:04

lot of people using multiple models but we're using multiple humans. our gangprompt system has a gang-grill mode — it asks several us for opinions to collectively arrive at a good design. → tweet link

@thdxr · 2026-06-24T01:56

it's so crazy ai is bad at your job but good at everyone else's job → tweet link

@TheAhmadOsman · 2026-06-24T01:43

I have never been more confident that Local and Opensource AI are going to win. LFG. → tweet link

@TheAhmadOsman · 2026-06-24T09:33

Qwen team, are you planning on releasing an opensource Qwen 3.7 model or should we just call your opensource contributions at this point? → tweet link

Hugging Face & Open-Source Ecosystem

@victormustar · 2026-06-24T08:16

At Hugging Face we've been building our own agent that we use via Slack (Moon Bot). Honestly, building your own is quite simple… any model you want, fully customizable, your data never leaves your infra, every session auditable. → tweet link

@victormustar · 2026-06-23T21:38

Cool model alert: futo-swipe 😍 a tiny set of CNNs that decode swipe gestures into text, fully on-device. the encoder is 635K params/2.65MB and works on ANY keyboard layout! → tweet link

@victormustar · 2026-06-24T07:46

some genius invented a kebab benchmark for llms → tweet link

@victormustar · 2026-06-23T19:34

wow first drop of Krea 2 Turbo LoRAs look very good 🚀 → tweet link

Developer Tools & Infrastructure

@LinusEkenstam · 2026-06-24T17:15

FIGMA MOTION 🔥 This is crazy, this was all done inside Figma, and you can export to code, mp4, webm, GIF and more coming. Figma is absolutely crushing it. → tweet link

@LinusEkenstam · 2026-06-24T16:01

CONFIG 2026 — LIVE from SF. I'll be covering all the latest drops from Figma in real time. → tweet link

@ASalvadorini · 2026-06-24T06:24

Hey @FlutterDev any chance you can make all the lint rules more agent friendly? My agent would like to be able to retrieve them all in a simpler format… → tweet link

@jezell · 2026-06-24T04:46 (RT @criccomini)

SlateDB now has distributed compaction AND sub compactions on main. New release incoming. → tweet link

@badlogicgames · 2026-06-23T22:08

People of Processing. Gramps made another mistake and broke forking company custom env fields in models.json… Update to 0.80.2. Sorry. → tweet link

@kunchenguid · 2026-06-23T19:54

has anyone been running claude/codex in github actions while still using your subscription tokens? that would basically become free cloud sandboxes for open source projects… → tweet link

@kunchenguid · 2026-06-23T23:37 (RT @ssbrouhard)

no-mistakes by @kunchenguid just blocked a merge over a privacy leak I never would've caught. New feature would've sent private data to a 3rd party. → tweet link

Big Tech & Talent

@kunchenguid · 2026-06-24T04:36

the person who created google workspace CLI got fired by google… Justin was DevRel — not on the core product team… from the workspace leadership's POV Justin was a random guy popping out of nowhere… you can't just ship things, even if they are good. → tweet link

@LinusEkenstam · 2026-06-24T04:44

Biggest miss from Google firing @JPoehnelt over this. Also heard @addyosmani left google, combined with other high profile exits last week. Google really need to rethink its talent retention mechanics. → tweet link

@steipete · 2026-06-24T01:31

Google fired the guy that made the google workspace cli, because he made the google workspace cli. Lucky me, Google can't fire me. → tweet link

@TheAhmadOsman · 2026-06-24T16:35

The economics of scale folks will keep repeating it as they worship their cloud providers until they get rug pulled hard and even then they'll be in denial about it → tweet link

Hardware & Open Hardware

@FrameworkPuter · 2026-06-24T00:33

Since LPCAMM2 is still hard to find retail options for, we're updating our Intel Core Ultra Series 3 Mainboard pre-orders to enable the option to pre-order LPCAMM2 at the same time. → tweet link

@FrameworkPuter · 2026-06-23T22:45 (RT @jlcjak)

FRAMEOSCOPE IS HERE — A fully open source oscilloscope(+FPGA) module for framework laptops! → tweet link

Miscellaneous Tech

@thdxr · 2026-06-24T12:18

nyc bans waymo and all of tech flips to moralizing. when will we realize that our industry is fucking terrible at getting the world excited about the future. → tweet link

@sudoingX · 2026-06-23T20:28

grok build is cracked at 3d and physics, so so so underrated. you iterate with it in plain english and it just keeps building on what's there. → tweet link

@levelsio · 2026-06-24T18:31

The craziest thing about Quake is how many modern FPS games originate from it somehow. Valve's Source engine was a modified Quake 1 engine… The entire Call of Duty and Warzone series originates from a modified Quake 3 engine. → tweet link

@badlogicgames · 2026-06-23T22:11

This is what working on SaaS for many years does to you. Barely any contact with your users on a minute to minute basis. No such bliss for desktop software. → tweet link

@MengTo · 2026-06-24T15:46 (RT)

I recorded a 22-min tutorial on how to avoid AI slop for your landing pages → tweet link

@thdxr · 2026-06-24T17:24

there's something weird about everyone saying how productive their token usage is but being entirely dependent on subscriptions and completely incapable of paying per token. is it providing returns or not? → tweet link