Executive Summary

The past 24 hours in the tech and AI landscape have been dominated by strong pushes toward open-source AI and local model deployment, directly challenging the centralized API subscription models of major closed-source labs. On the hardware and infrastructure front, new quantization techniques and Apple's M5 Max chips are making local, high-context agentic workflows increasingly viable, even as cloud GPU rental prices surge by over 200%. In software development, AI coding agents are proliferating—with Grok Build, Claude Code, and Cursor competing fiercely—while developers are actively optimizing tooling, replacing heavy dependencies with Rust/WebAssembly alternatives, and confronting the practical limits and "psychosis" surrounding AI's current capabilities. Meanwhile, significant AI research breakthroughs, such as DeepMind solving open Erdős problems using LLM-Lean agents, highlight the rapid maturation of automated reasoning.

Key Events

• A major manifesto advocating for open-source AI as "civilizational infrastructure" goes viral, warning against cognitive subscription economies controlled by a few closed labs → link
• DeepMind uses LLM-Lean autonomous agents to solve 9 open Erdős problems, showcasing a breakthrough in AI mathematical reasoning → link
• DeepSeek V4 Flash returns to the Nous Portal for free use in Hermes Agent → link
• Qwen 3.5 27B in NVFP4 quantization runs full context in under 20GB VRAM, enabling multiple fast local agents on a single RTX PRO 6000 → link
• McKinsey is reportedly "under pressure from clients" to shift from billable hours to AI-driven value pricing, signaling a major disruption in traditional professional services → link
• GPU rental prices have surged by 200%+ in the last 5 months, adding cost pressure to cloud-dependent AI development → link
• Apple's MacBook Pro M5 Max shows unprecedented speeds running local models like Qwen 3.6 35B → link

Analysis

Patterns & Trends: There is a clear bifurcation in the AI ecosystem: closed-source labs are pushing expensive, potentially "nerfed" subscription models, while the open-source community is rapidly advancing quantization (NVFP4) and local deployment to bypass cloud costs. The 200% spike in GPU rental prices only accelerates this pivot toward local compute and efficient inference (e.g., Qwen TokenSpeedMLA). In software engineering, "AI psychosis" is being called out by developers who realize that while AI accelerates coding, the overall speed of project delivery still feels bottlenecked by legacy integration and last-mile realities.

Escalation/De-escalation: Tensions are escalating between open-source advocates and closed API providers, framed explicitly as a fight for "operational freedom." Simultaneously, the traditional consultancies (McKinsey) are being forced into de-escalation, pivoting away from billable-hour models as AI startups directly target their business structure.

What to Watch Next: Watch for increased adoption of Rust/WebAssembly in AI toolchains to reduce bloat (following the OpenClaw dependency purge), further development of agentic evals to cut through the hype of LLM outputs, and the ramp-up of Claude Code's upcoming /workflows feature which could redefine standard developer pipelines.

Tweet Feed

AI Models & Hardware Performance

@TheAhmadOsman · 2026-05-25T09:21

Qwen 3.5 27B in NVFP4 w/ full context taking less than 20GB VRAM

You can basically run like 5 agents w/ full context on a single RTX PRO 6000 like this, and they'd be so fast

Tell me I didn't tell you this was gonna happen https://t.co/ppg6ZGQMi4 → tweet link

@Ex0byt · 2026-05-24T19:15

Qwen3.7 + Qwen TokenSpeedMLA is going to be insane for local long horizon agentic workflows.

(Blog: https://t.co/dghB7RtgDW) (Code: https://t.co/7SRr3MlgSH) → tweet link

@Teknium · 2026-05-25T04:44

DeepSeek V4 Flash IS BACK on Nous Portal for FREE for use in Hermes Agent!

Check it out at https://t.co/tMAQFkegul → tweet link

@RydMike · 2026-05-25T18:00

RT @bridgemindai: MacBook Pro M5 Max is fully set up and running local models.

I have never seen speeds this fast.

Qwen 3.6 35B and Gemm… → tweet link

@TheAhmadOsman · 2026-05-25T08:48

GPU rental prices went up by 200%+ in the last 5 months btw → tweet link

@gospaceport · 2026-05-25T14:16

Charts like this are how European GPU wars start. → tweet link

@gdb · 2026-05-25T03:22

GPT-5.5 Pro for fact checking: → tweet link

AI Agents & Developer Tools

@sudoingX · 2026-05-25T15:29

i keep coming back to grok build. theres something about having grok in the terminal that just hits different.

ive been playing with it so much that it built its own memory system and registered itself as my official peer agent. full scaffold. all matched to my existing agent convention.

native x fetch is what changes everything. live x search and context inside the terminal next to my code, my agents, my benchmarks. ask what the room is saying and the model answers from the live feed.

first coding agent that feels native to how i actually work. → tweet link

@sudoingX · 2026-05-25T13:28

claude code monthly subscribers are being served a nerfed version of opus 4.7. i'm convinced now after using cursor on same model.

anyone serious about agentic software development should be on cursor right now.

cursor on opus 4.7 max consistently outperforms claude code on opus 4.7 max.

if it's the same model, why does one feel like the real thing and the other feel like a shit? → tweet link

@sudoingX · 2026-05-25T06:36

someone at xai actually uses grok build on a phone. you can tell.

look at this. grok build TUI on termius mobile, with the keyboard up. last prompt i sent is pinned at the top. grok's output is in the middle.

current input box sits right above the keyboard. i can see what i asked, what it answered, what i'm about to type, all on a phone screen, while the keyboard takes half the viewport.

zero flickering. scrolls clean. works in tmux. works on termius. works on a laptop. works on a phone. xai actually optimized this TUI for how devs USE it, not just how it looks on a 4k monitor.

most coding agents pretend mobile doesn't exist. this one made mobile a real workflow not an afterthought.

small detail. the kind that means the team is paying attention. → tweet link

@Teknium · 2026-05-25T12:18

Some new improvements to performance just went in.

Python gets a bad wrap for performance but we aint looking to shabby against a trillion dollar co's rust codebase, beating codex at most multi-turn tasks we benchmarked (mt stands for multiturn)

PR: https://t.co/nNJz3EPAPA https://t.co/6z7vXpLA9X → tweet link

@kunchenguid · 2026-05-25T17:05

ugh.. it’s a really bad advice to tell people to keep their laptop lid open as a solution to keep agents running

you don’t need tmux. you literally just need to run a single command:

sudo pmset -a disablesleep 1

you can even ask your agent to help you make a script to run this as one-off instead of being a persistent setting

the problem with keeping the lid open is obvious: - holding a half open laptop is very uncomfortable - you may accidentally close it - other people can grab your laptop and there’s no password protection → tweet link

@steipete · 2026-05-25T14:27

Folks: when you write skills, ask your agent to be token efficient, relax grammer. I see too many skills that write books in the skill description, and all that crap is loaded into every context.

I wrote a skill that finds the worst offenders. https://t.co/kfaaJpxMXE → tweet link

@steipete · 2026-05-25T12:10

New pet peeve: cli's that install new skills onto my system without asking. → tweet link

@victormustar · 2026-05-25T13:34

I now literally spend my days having AI agents read and analyze my other agents sessions... https://t.co/QfRyljcCtd → tweet link

@badlogicgames · 2026-05-25T09:01

trying it outdoors without its robot body, and it's more useful than flicker or gerpertee mobile apps + voice mode. i can also give it tools like spotify, email, etc. so i actually get working android auto.

what i'm saying is: take control of your LLM apps. it's actually not hard to make something better than the labs. → tweet link

@badlogicgames · 2026-05-25T01:38

introducing pipi, the shitty robot.

brain lives on my laptop, sensors/UI live on the mounted phone. time to completion: 24h (minus sleep, knight festival, lunch, dinner, and play)

built with https://t.co/oUoqqL9hAp https://t.co/DiLLLuecwN → tweet link

@MengTo · 2026-05-25T12:49

I recorded a 27-min tutorial on how to prompt beautiful landing pages with animated images 2.0 https://t.co/aqleYl6ALN → tweet link

AI Research & Open Source

@TheAhmadOsman · 2026-05-25T18:33

OF CRUCIAL IMPORTANCE - PLEASE READ

If intelligence becomes something people can only rent from a few closed institutions, the public does not just lose software freedom. It loses operational freedom.

AI is a civilizational infrastructure for work, education, science, software, creativity, public services, and national capacity.

This civilizational infrastructure must not become rented access through closed APIs, remote platforms, shifting terms, opaque moderation, and prices set by a handful of companies.

The ability to study, build, repair, deploy, audit, adapt, teach, preserve, and run intelligence systems without asking permission is of EXISTENTIAL importance.

With OpenAI, Anthropic, and a handful of other players controlling the models, this civilizational infrastructure risks becoming a subscription economy for cognition.

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.

America should not fall behind on the freedom to run, inspect, modify, benchmark, teach, and preserve intelligence infrastructure. The practical posture is American capacity with global open standards. → tweet link

@TheAhmadOsman · 2026-05-25T03:29

DROP EVERYTHING

The ultimate step-by-step projects roadmap for BECOMING an AI Researcher is now available online to read FOR FREE

Covers building

• Tokenizers / embeddings
• Positional methods
• Attention / multi-head attention
• Transformer blocks
• Training loops / objectives
• Sampling dashboards
• Speculative decoding
• KV cache / MQA / GQA / MLA
• Long context
• FlashAttention / hardware budgets
• MoE routers
• State-space / diffusion LMs
• Data pipelines / synthetic data
• Scaling laws
• SFT / DPO / RLHF / GRPO / RLVR
• Quantization
• Serving systems
• Evaluation harnesses
• RAG / tools / agents
• Multimodal adapters
• Interpretability / safety
• Full capstone model system

The loop for every project

• Build it
• Plot it
• Break it
• Explain it
• Ship the artifact

You should read this, and if you cannot now then you most definitely wanna bookmark it for later

DM me when you're working at a frontier lab → tweet link

@TheAhmadOsman · 2026-05-24T20:52

I’ve acquired https://t.co/uczzOF3bV8 & https://t.co/zlFpCV6H3t (alongside a few other related domains) recently as well

Open science and sharing our knowledge is the only way forward

Opensource / Local AI FTW → tweet link

@RealGeneKim · 2026-05-25T14:11

RT @prz_chojecki: Another 9 open Erdos problems solved, this time by DeepMind team.

Interesting loop of LLM - Lean agents working autonomo… → tweet link

@jsuarez · 2026-05-25T18:36

Reinforcement learning research with Joseph Suarez https://t.co/ZKkvUTULG9 → tweet link

@badlogicgames · 2026-05-25T17:54

RT @Vtrivedy10: nice write up from the HuggingFace folks aggregating works on defining agents, harnesses, environments, RL, etc. The more… → tweet link

@juliarturc · 2026-05-25T16:42

"World models" is one of the buzziest yet ambiguous terms in AI right now. I started this video with many questions: - How are they different from video generation? - Can they do more than AI slop? - Can LeCun be trusted given that he wears knee-high white socks?

Many thanks to @tjgalda and @NVIDIAAI for helping me answer (most) of these questions! → tweet link

@Teknium · 2026-05-25T18:36

RT @NousResearch: Join the team on Wednesday for another Hermes Agent Jam! https://t.co/WWa56dci82 → tweet link

@Teknium · 2026-05-25T01:34

Welcome to the Hermes Agent crew!

Let me know if you run into any problems or have suggestions! → tweet link

@Teknium · 2026-05-25T00:41

RT @JulianGoldieSEO: HERMES JUST FIXED THE BIGGEST PROBLEM WITH BROWSER AGENTS

Most AI agents still click around websites like confused in… → tweet link

Software Development & Programming

@steipete · 2026-05-25T14:44

OpenClaw's dependency purge continues. Killed Sharp and Jimp. Replaced it with photon, a small WebAssembly that runs compiled Rust for image processing. 2MB vs 140MB. https://t.co/tSimX2GKwP → tweet link

@ASalvadorini · 2026-05-25T17:37

Pro tip: pre-compile your SVGs for better performance 🔥 👇👇👇

Flutter #Flutterdev

→ tweet link

@ASalvadorini · 2026-05-25T03:28

This is not by luck. It's the result of years of solid work by the Flutter team, the intrinsic easiness and performance of Flutter, combined with the AI era.

Prediction: ⬇️⬇️⬇️

Flutter #Flutterdev

→ tweet link

@badlogicgames · 2026-05-24T23:39

wow, someone should by the uv folks, it makes python bearable. → tweet link

@RealGeneKim · 2026-05-25T14:20

RT @crvvdev: Did you literally know that Windows has something called Warbird that literally executes encrypted shellcode on your computer?… → tweet link

@uwteam · 2026-05-25T18:14

Gdy tworzysz linka do profilu/posta w social mediach i umieszczasz go na dowolnej stronie, to pojawia się pewien problem - po kliknięciu go na smartfonie, przeważnie otwiera się wbudowana przeglądarka WWW, a nie zainstalowana aplikacja. Ten serwis rozwiązuje ten problem: https://t.co/VzaAJDiCeD

Jest pełno takich SaaS-ów ($), ale to jest darmowe, opensource i moje 😏 → tweet link

Tech Industry & Commentary

@TrungTPhan · 2026-05-25T14:53

RT @bearlyai: McKinsey is “under pressure from clients” to change its business model due to AI.

Instead of tying fees to hours worked —AI… → tweet link

@TrungTPhan · 2026-05-25T17:28

RT @bearlyai: more and more AI startups marketing against the traditional professional services billable business model https://t.co/oFpYW1… → tweet link

@levelsio · 2026-05-25T18:49

I'm pretty sure by now this entire thing is perfomance art

Everyone in tech is falling for it

I also doubt he actually raised money

Even the name in reverse is just AI Slop

Kinda Sacha Baron Cohen style but for startups! 👏 → tweet link

@RealGeneKim · 2026-05-25T14:09

RT @DanielMiessler: Claude Code is about to release a feature called /workflows that I think will be extremely significant.

Especially fo… → tweet link

@RealGeneKim · 2026-05-25T14:09

RT @levie: CEOs are uniquely prone to AI psychosis because they’re sufficiently distant from the last mile of work that still has to happen… → tweet link

@LinusEkenstam · 2026-05-24T20:53

A platform starts to police what gets posted. It stops being the internet. It becomes an intranet.

The internet works because it's the internet. Circular. Intentional.

Look at the art world. Artists rarely sell direct to consumer. They use gallerists. They get solo shows because they have talent management or representation. Strange model. But it mirrors the internet.

Plenty of verticals do.

When platforms become the police, censoring or dictating what people post and how they behave, it never ends well. We saw it with Twitter and politics. That's probably why Elon got mad in the first place and bought the joint.

The scum find a way. The people at the bottom don't care. The middle gets squeezed.

Social media is broken on many levels. Trying to fix it is worth cheering for. But the internet stops being fun and random the moment you start to police it.

Look at how people actually use it. 90 to 99% of users are passive. They lurk. That leaves 1 to 10% creating and contributing.

Inside that small group, there are types.

There's the curator. The one who filters signal from noise. Most good creators do this well. They learn what works. Some bring their own voice and opinions. Some just curate.

Then there's the artisan. The one who crafts, molds and perfects their own stories. These tend to live in a niche. Unlike the curator, they're stuck there. The audience boxes them in. The algorithm boxes them in.

Some of the most influential creators in the world are one or the other. You can have reach and impact while staying in a niche. @Casey is a vocal pro-New York visual storyteller in the vlog sphere. @TheB1M is the number one big construction channel on YouTube. Both huge. Both influential. Both stay in their lane. There's a roof on their reach and that's fine.

Both curate. Both try to understand what works. How to tell a story. But they tell those stories themselves. The largest bridge in the world. The reason behind writer's block while being pro New York.

Then there's the opposite. The one-take reposter with the occasional shill or personal win. The biggest one is the platform owner himself, @elonmusk

He's been active on Twitter/X for a long time. Uses it like a toilet diary. Random shitposts with commentary. Sometimes just videos or photos with no context.

He shills his companies. He shills his friends. He dunks on competitors. He tries to sway elections. He is the definition of a digital curator with a personal twist. Or sometimes the lack of one.

Then there are the programmatic reposters. Hundreds of accounts pushing one classic movie clip every 20 minutes, 24/7, 365. Stolen IP. Massive reach. The studios and IP holders don't complain. It gives them free eyeballs.

This dynamic, broad, nuanced mix is what makes a platform like X interesting.

In between every example sits more nuance. More complexity. More openness. More opinion. The collective hive mind of the internet.

Modern algorithms have good intent. But they break this dynamism. They erode the core that makes the platform vibrant. It gets worse when the platform itself steps in to intervene.

Good intentions. Bad outcome.

It's no longer the open internet. It's the xAI intranet. → tweet link

@thdxr · 2026-05-25T03:31

think back to projects you've worked on in the past

it's hard not to imagine they'd have been completed way faster now that we have ai

but everything still feels as slow and as difficult as ever → tweet link

@nummanali · 2026-05-24T23:34

TLDR: I’m so done building for the sake of it that I’m taking a break

In September 2025 I made my first major OSS project - OpenSkills, it’s now at 10K+ stars

It gave me a sense of purpose and meaning, because I was solving a problem for people around the world

There have a been a few other projects ie Codex OAuth for OpenCode, CC Mirror and Wrapped projects

As models grew better, I ended up trying more ambitious ideas such as Local LLM inference on Macs, Agent orchestration, and ideas such as Notion for agents etc

These larger ideas, tens of them, grew to the point that they felt like vanity projects over purpose and largely in part due to not properly vindicating the code through my eyes

What I mean, is that it was built with excitement over love and care

I’ve been a humble user and advocate of agentic engineering since early 2024, so I say this with heaviness that I have deliberately - slowed down

It simply is not worth the effort or wasted mental fatigue trying to build creations of ineptitude

I’ve shelved the remnants of them, for now, and instead decided to observe the landscape as it shifts week by week

I only use one terminal now, with one task at a time

I feel I am at a transitional phase where I am taking deeper interest in how the tokens are generated and the skews in outputs occur based on the multitude of factors within a LLMs architecture

It doesn’t mean I’m against the premise of the latest innovations and methodologies to maximise LLM backed outcomes, it simply that the value of over engineering the set up and implementation is a narrow percentage return at this stage

We need better signals on what derives good outcomes from LLMs, not the majority he says/she says we hear everyday, it’s a form of eval but one that likely has more nuanced elements to it - think ARC AGI but in profession functions in vertical slice

I believe I’m dreaming of Agentic Evals that simulate unpredictable scenarios like the real world of software engineering throws at you → tweet link

@TrungTPhan · 2026-05-25T01:41

RT @lennysan: Automation is a lie. CLIs are over. The SaaSpocalypse is dumb.

A year ago @danshipper came on the podcast to predict where A… → tweet link

@victormustar · 2026-05-25T14:34

Question for robotics people: how far are we from a ~75cm toy robot that can walk and move around okay-ish (bonus points for claw hands) AND costs around $1 or 2k?

Feels like we're technically close, but I haven't seen anything. What am I missing? hardware cost must be the bottleneck? → tweet link

@TrungTPhan · 2026-05-25T15:12

Pope Leo working XIV working with Anthropic on AI is the biggest tech and Vatican crossover since Steve Jobs studied Catholic Chuch to make Apple’s org chart:

▫️“We've observed that the oldest and largest organization in the world has only four layers of management. 

That's the Catholic Church. And, uh, five if you count the highest order I suppose.

So, we see no reason why we need over four layers of management. Indeed, we have usually about three. That’s the president. Maybe, we have a division manager and then maybe under that a marketing or engineering manager.

And that's really about it. So, that's what we're trying to do. […]

We hire people to tell us what to do. [They] come back and tell us how much it's going to cost and go do it. We've got an incredible group of entrepreneurs…[these] people are very independent thinkers and what they really want is…the environment where they don't have to convince 30 other people that it's the right thing to do. […]

We really make an attempt to do that and I guess our feeling is that the day that somebody working in Apple decides that they can't make a difference anymore, is the day we've lost.”▫️ → tweet link