← Tech / AI / IT Monitor Index Tech / AI Generated 2026-06-25 19:31 UTC

Tech / AI / IT Monitor

June 25, 2026 · Based on tweets from the last 24 hours · 134 tweets analyzed · model: ollama-cloud/glm-5.1:cloud

Daily Intelligence Briefing: Tech / AI / IT Monitor

Date: 2026-06-25

Executive Summary

The past 24 hours were dominated by the accelerating local AI movement, with developers demonstrating DeepSeek V4 running on single DGX Spark units via model pruning, and multiple open-source agents (Hermes, Ornith, Fable) gaining new capabilities. Hardware supply-chain pressure intensified as Micron posted record earnings driven by HBM demand, Apple raised MacBook/iPad prices citing memory costs, and commentators noted that local AI ambitions remain bottlenecked by RAM scarcity. Meanwhile, MCP came under criticism for architectural overhead, and early signs of "Fable 5" as a Claude Code subscription tier were discovered in binary diffs. Open-source AI models and tooling continue to outpace closed ecosystems in community momentum.

Key Events

Analysis

Local AI is moving from aspiration to demonstration, but hardware remains the bottleneck. The DeepSeek V4 on DGX Spark benchmarks are genuinely significant — showing that with pruning (REAP), alternative inference engines (vllm + speculative decoding), and quantization, frontier models can run on single machines. However, @jezell's counterpoint that local AI is a "pipedream" due to RAM supply constraints, and @levelsio's admission that a single RTX 5090 is "painfully slow and useless" for local LLMs, underscores that this is still an elite-hardware game. The memory supply chain (Micron HBM, Apple's price hikes, RAM 4x-ing in cost) will determine how fast this democratizes.

Agent ecosystem fragmentation is accelerating. Hermes, Fable, Ornith, Claude Code's embedded agents, and Meta's auto-research agent are all competing for the "default agentic layer" with different philosophies. MCP's architectural limitations are already creating friction. The next 60 days will likely see consolidation or de facto standard emergence.

"Vibe coding" backlash is forming. @thdxr's metric (rg -o 'isRecord' | wc -l) and @hnasr's prediction that AI will produce "bloated products" that need expert troubleshooters signal a growing awareness that AI-accelerated development has quality costs. This aligns with @mipsytipsy's argument that "AI demands more engineering discipline, not less."

What to watch: Fable 5 pricing/release; whether REAP pruning gets adopted beyond DGX Spark; Micron's HBM guidance next quarter as a bubble indicator; MCP spec evolution or fragmentation.

Tweet Feed

AI Models & Research

@gdb · 2026-06-25T17:37

Agents are being adopted very quickly and accelerating work. How this looks across OpenAI itself: https://t.co/pQEiAUWuDa → tweet

@sudoingX · 2026-06-25T14:43

my dgx spark is sitting on my desk running a 180 billion parameter deepseek v4. just one machine anon not a cluster. a model this size normally doesn't fit on a single box. what makes it possible is @0xSero's REAP, he pruned deepseek v4 from 284B down to 180B so it actually loads on a single spark with room to run. [...] nearly double, just from the engine plus specdecode trading the spark's idle compute for speed. → tweet

@sudoingX · 2026-06-25T11:47

i just got DeepSeek V4 Flash running on a single dgx spark, the 180B REAP'd down to 82GB so it actually fits with context headroom, on antirez's ds4 engine. i'm seeing ~65 tok/s prefill and ~13 tok/s generation (Q3, single stream). → tweet

@victormustar · 2026-06-25T17:01

Cool: if you have a Hugging Face account you have enough free credits to ask GLM-5.2 to build your website on HuggingChat (it will use Exa to do some research about you) and you can deploy it for free in a static HF Space. Also this model has good taste 🤌 → tweet

@victormustar · 2026-06-25T14:30

not going to lie Fable orchestrating Sonnet 5 subagents could be something... 👀 → tweet

@Ex0byt · 2026-06-25T16:32

Currently in the inference wars... Autobots, roll out! → tweet

@Ex0byt · 2026-06-25T16:03

new skill unlock for your auto-research agent from ai@meta → tweet

@Ex0byt · 2026-06-24T22:01

Awesome work from the FireworksAI_HQ team. Not an easy feat to accomplish on a model this size. → tweet

@jsuarez · 2026-06-24T19:01

Reinforcement learning research with Joseph Suarez → tweet

Open-Source AI & Local Inference

@TheAhmadOsman · 2026-06-24T20:26

We're gonna make sure Opensource and Local AI win. Watch us make that the default. → tweet

@sudoingX · 2026-06-25T16:25

almost a year of building. days, nights, weekends, holidays. all of it pointing at one thing. it's almost ready. and when it lands, local ai stops being intimidating for everyone who's been waiting to start. → tweet

@jezell · 2026-06-25T17:19

This is why local ai is a pipedream. Get in line behind everyone else trying to buy a lot more RAM than you. Maybe in 10 years... → tweet

@levelsio · 2026-06-25T18:53

And yes I've tried local LLMs but with just 1x RTX 5090 it's painfully slow and useless → tweet

@victormustar · 2026-06-25T14:29

RT @ClementDelangue: Qualcomm! → tweet

@victormustar · 2026-06-25T15:11

RT @huggingface: Welcome to Open Source AI: Run Your Own Models Locally → tweet

@victormustar · 2026-06-25T14:49

RT @ornith_: Aloha! 🌺 Meet Ornith-1.0, a family of open-source LLMs specialized for agentic coding. → tweet

@victormustar · 2026-06-25T13:56

RT @mishig25: In HF GGUF section of models, we are emphasizing MTP heads with its own sign 𝗠𝗧𝗣 → tweet

@victormustar · 2026-06-25T13:38

RT @Thom_Wolf: To all the newcomers excited to try Opus 4.8-level models at home: welcome to OpenWeightLand! → tweet

@TheAhmadOsman · 2026-06-24T21:31

Best use of Codex Cli is to ask it to teach you how to get LLMs running locally btw → tweet

@TheAhmadOsman · 2026-06-24T23:14

We're eating good boys with a DGX Station tonight → tweet

@tinygrad · 2026-06-24T21:43

RIP, no future in Mojo now. Qualcomm is not a good steward for open source software. Modular raised too much money, should have stayed leaner to have a chance at winning. → tweet

Agent Tooling (Hermes, Claude Code, Fable)

@sudoingX · 2026-06-24T20:42

fable 5 incoming! pulled the actual 2.1.190 binary of claude code and diffed it against 2.1.187 to see what's really in there. "you've used your included fable 5 usage for this week" is a brand new string, didn't exist in the last version. fable 5 getting a weekly allotment in the sub would go hard. but the rumor oversold it. → tweet

@Teknium · 2026-06-25T01:08

How is /learn going for all of you? Any issues or improvement ideas anyone has? → tweet

@Teknium · 2026-06-24T20:38

Great extra context for the new /learn capability in Hermes Agent! → tweet

@Teknium · 2026-06-25T16:21

RT @honchodotdev: OAuth with Hermes Agent is now live! @NousResearch Connect to Honcho in 1-click from Hermes Desktop and the CLI → tweet

@Teknium · 2026-06-25T16:13

RT @NousResearch: Choose your own → tweet

@Teknium · 2026-06-24T20:37

RT @NousResearch: Sometimes you just need a dose of fresh inspiration but your agent doesn't get the vibe. The creative-ideation skill analyzes... → tweet

@Teknium · 2026-06-25T07:28

RT @tobi: This is so good. I gave my hermes a budget and now I'm getting 'gifts' in the mail. Highly recommended. → tweet

@kunchenguid · 2026-06-24T21:03

omg /no-mistakes is trending on github today! and sitting above hermes?! → tweet

@kunchenguid · 2026-06-24T19:56

claude down. 3rd day in a row → tweet

Developer Tools & Frameworks

@thdxr · 2026-06-24T23:42

mcp is so tied to the idea of 1 process = 1 session. there's some things in the spec that force you to spawn a new MCP server for every active session you have. ridiculous overhead - this is why i kept saying it's too early to standardize → tweet

@thdxr · 2026-06-24T23:34

i found a really good way to measure how much a codebase is suffering from vibe coding: rg -o 'isRecord' . | wc -l → tweet

@thdxr · 2026-06-24T23:58

it's worth deeply studying why no framework dethroned react. it's completely misunderstood and it's why every prediction you see by programmers tends to be wrong → tweet

@thdxr · 2026-06-25T14:18

people that like verbose apis: it's more explicit for the agent. people that like terse apis: it's more token efficient for the agent → tweet

@ASalvadorini · 2026-06-25T07:45

Minimal 3.0.0 is out 🔥 It has a very small breaking change which shouldn't effect anyone + I introduced a release agent 🚀, so I can ease my flow. This release was all done by the agent itself 🙏😇 → tweet

@ASalvadorini · 2026-06-25T12:18

Spoiler: I really love release agents 🔥 this is the fourth I write in a month. If you don't have one already, I strongly recommend you to write one, check the Minimal one as a reference → tweet

@ASalvadorini · 2026-06-25T06:01

Can some iOS guru explain me like I'm 5 years old why the moment I update Xcode I'm not able to deploy into the simulator (which has been working totally fine) again until I download the new iOS? → tweet

@ASalvadorini · 2026-06-25T07:00

A new Flutter version 3.44.4 is out 🔥 → tweet

@cooltechtipz · 2026-06-25T15:42

MCP learning guide → tweet

@cooltechtipz · 2026-06-25T11:33

The AI skills developers should learn → tweet

@cooltechtipz · 2026-06-25T05:46

Build, deploy, monitor, scale with LLMOps. → tweet

@cooltechtipz · 2026-06-25T02:09

From APIs to AI agents → tweet

@jezell · 2026-06-24T21:21

Until @figma is shipping a Figma runtime like @rive_app, I feel like everything they announce is pointless. → tweet

@MengTo · 2026-06-25T13:44

I recorded a 10-min tutorial on how to use Figma Motion and shaders. → tweet

@MengTo · 2026-06-25T02:21

If anyone can solve motion design, it's Figma → tweet

@LinusEkenstam · 2026-06-25T01:02

I'm obsessed with Figma → tweet

@LinusEkenstam · 2026-06-24T19:14

So much of my last 13 years have been spent inside Figma. There are thousands of people that made it possible. But without Dylan none of this would be happening. → tweet

@badlogicgames · 2026-06-24T19:37

good stuff, @github → tweet

Hardware & Memory Supply Chain

@jezell · 2026-06-25T16:58

It's been a while since Apple was the little dog in the memory supply chain. → tweet

@TrungTPhan · 2026-06-25T14:49

Based on how much Apple is raising prices for Macbooks and iPads due to higher memory prices, the upcoming iPhone Fold legit might cost $2,999. → tweet

@FrameworkPuter · 2026-06-25T14:25

In response to Apple's price increases today, we've lowered the price of some Framework Laptop 13 Pro configurations. We were able to source and qualify Gen 5 SSDs from ADATA that are both faster and cheaper, and now offer them on DIY Edition! → tweet

@levelsio · 2026-06-25T18:01

My gaming PC I built last year to play Flight Simulator 2024 has almost doubled in value now. A year ago I paid €6,562 for it, now it's worth €12,073 (+84%). The GPU 2x'd, RAM 4x'd, and SSD up a bit → tweet

@FinansowyUmysl · 2026-06-25T08:43

Wczoraj $MU opublikował rekordowe wyniki! Rynek oszalał. Jeszcze lepsze są prognozy na kolejny kwartał - EPS: 31$. Główny powód to: eksplozja popytu na pamięć HBM do serwerów AI. → tweet

AI Industry Commentary

@TheAhmadOsman · 2026-06-25T18:51

Karpathy has had a massive aura loss after joining Anthropic, really sad to see. → tweet

@TheAhmadOsman · 2026-06-25T09:25

Without naming names, that Slack integration from the famous AI fearmongering and rugpulling company is made to steal your business intellectual property and create themselves new moats. → tweet

@hnasr · 2026-06-25T14:02

In the end, AI will produce more bloated products for us to troubleshoot and fix. If you are a good software troubleshooter who understands the fundamentals, you will be on high demand. And your skillsets are deadly. → tweet

@mipsytipsy · 2026-06-24T20:17

i said i would write a followup piece on AI and ethics, and I have. We do not get to choose to live in a pure world. But we get to choose whether to help shape the compromised world we already live in. → tweet

@sudoingX · 2026-06-24T20:33

let me make something clear since it keeps coming up. you cannot buy my opinion. not with access, not with an intro to your team, not with a friendly DM after i call your product bloat. [...] i test it, i tell you what the data says, and the data doesn't have a price. → tweet

@levelsio · 2026-06-25T18:55

A fun exercise is exporting your 5+ year of chat logs and putting it in Claude Code to do psychoanalysis of yourself → tweet

@louszbd · 2026-06-25T04:35

I spend 10 minutes a day trying a new app. https://t.co/6eRNDoBBSE is my favorite app today. [...] The thing I love about matrix is it helps you nail down what to build and why before it runs the whole thing. GLM-5.2 now works in it, and the experience is smooth. → tweet