Wafer made AMD GPUs competitive with NVIDIA chips for AI inference - at half the cost.
wafer: 🚨 BREAKING:
these engineers figured out how to serve GLM 5.2 on @AMD MI355X at 2626 tok/s/node and 213 tok/s single stream at over 2x lower cost than Blackwell
that's ~80% of B200 throughput at over 2x lower cost
full write-up in reply to see how

DOGE deletes itself on July 4th. It will be remembered as a hugely destructive failure.
Musk promised $2 trillion in savings.
What we got, by DOGE’s own unverified math, was $215 billion. Even that number was never proven.
The costs were atrocious. Nearly 140,000 federal workers pushed out. Veterans waiting longer for care. USAID dismantled while people died overseas as a direct result. 20-year-olds with Red Bulls and no college degrees rummaging through Americans’ most sensitive data.
Musk, the world’s richest man, a federal contractor with billions in government business, auditing the government that pays him, accountable to no voter anywhere. 
Even Musk now says he wouldn’t do it again. The guy who ran it calls it a mistake.
Government reform was there for the taking. Both parties know federal technology is outdated. Both parties know waste exists. Fixing it required serious people working with career civil servants who know where the problems actually are. Instead we got a chainsaw, a photo op, and years of lost expertise we now have to rebuild.
Now Trump’s budget director says there will be no final accounting or receipts. They promised transparency and they’re closing up shop in the dark.
They’re counting on us all to forget.
https://www.eenews.net/articles/doge-self-deletes-on-july-4th-the-grand-experiment-fell-apart-long-before-that/

happy was able to see you!
@aiDotEngineer means a lot to me too, and so many others. Not only the densest signal conference, but also the best place to make new friends for life.
The community is so lucky to have @swyx; he is truly gifted at connecting people.
Alex Volkov @ AI Engineer: I'm back home, exhausted but very emotionally fulfilled.
@aiDotEngineer means a lot to me, it means seeing folks I've met a few years ago now, and chatting about what's new in their life and how they are using AI.
It means meeting new people and adding them to that circle, it

This is actually a very good (and short and simple!) guide.
A bunch of neat tricks in there too.
Matt Shumer: Here's my in-depth guide to getting the most out of Claude Fable 5, so you can build things as insane as my demos below.
https://simplemarkdowneditor.com/pub/IbaCrTjLJT?key=uQOQ2NPO3TTUSXyYDjyLf

How? Instead of closed-source frontier labs running the same training runs in secret and siloes, open science and open-source AI allows them to mutualize spending and compute making them an order of magnitude more efficient!
Rand Group: 🇺🇸 The United States will spend $1 trillion on AI in 2027.
🇨🇳 China will spend $123 billion and still ship frontier models.
How?

My episode tomorrow is an inside look at how an OpenAI PM uses Codex for product work.
Rohan (Codex PM) walked through how to:
→ Use Image Gen to explore designs fast
→ Trigger Codex automations from Slack
→ Use one Codex thread to manage others
📌 Subscribe to get the full episode tomorrow: https://www.youtube.com/@PeterYangYT?sub_confirmation=1

On our 250th birthday, celebrating the contribution of immigrants and international collaboration
—46% of people with doctoral-level degrees working in US science and engineering fields are foreign-born
—41% of the science and engineering research published by US authors in 2024 included international collaborators
—20% of physicians working the the USA were born and educated abroad
@ACarnegieFdn and @TheLancet
https://www.thelancet.com/journals/lancet/article/PIIS0140-6736(26)01333-4/fulltext

Re Yet we still don't have level-5 self-driving cars, and certainly not self-serving cars that can learn to drive in a few hours of practice like any teenager.
We don't even have domestic robots that can do what 10-year olds can do the first time we ask them.
We don't even have robots that are nearly as smart as a house cat.
The G in AGI is nonsense.

Re @elonmusk @Devon_Eriksen_ If there were any lingering doubts that you and your friends were against the very principles of democracy, you just removed them.

Here's a cool piece of LLM lore: the original scaling laws were wrong due to a bug, which probably led to a lot of wasted compute on oversized undertrained models 🫣 (and that was before we even started properly accounting for inference cost!)
Diogo Almeida: http://x.com/i/article/2073276453131780096

spending the last week at @aidotengineer was awesome. too many great convos to cover them all, but jotted down some things that stood out:
- Lots of discussion around open source models. I spoke with consultants helping enterprise start to switching over low level tasks. Model routing companies and tools are seeing an uptick in usage and demand, and more innovation happening in routing (eg deterministic classifiers, etc).
- As an extension of this topic, local model room was packed. More people are running or at least trying to build fully local coding agents, primary reason seems to be cost (it’s people spending lots on tokens).
- They had a special area for “token billionaires” who are not people who have a billion tokens but spend a billion tokens… a week.
- Saw that OpenClaw was a topic on some talks, but in my private conversations (biased by who I spoke with), I spoke with many people who had switched to Hermes Agent, and very few active OpenClaw users.
- Fable came back during the conference (the strong Claude model that was temporarily banned by US govt). I heard people at the conference that it was butchered and not as good, but I also heard from people I trust that they didn’t notice a difference and that it’s great. It likely depends on what you’re using it for - my limited experience from this morning has been great.
- Getting coding agents to work more reliably, over longer horizons, on bigger code bases was a constant discussion and many popular talks and workshops on this. I was pleasantly surprised at how much of the content and discussions really felt like engineers sharing their own tips, not selling or marketing their product (though there was plenty of that too).
- The “company brain” topic came up a good amount as well. Distinct from “company knowledge/data”, this conversation was more around documenting how companies make decisions, not what data exists in their database.

Sent this straight to my OpenClaw to make a skill for prompting Fable.
Recommend you do the same.
Matt Shumer: Here's my in-depth guide to getting the most out of Claude Fable 5, so you can build things as insane as my demos below.
https://simplemarkdowneditor.com/pub/IbaCrTjLJT?key=uQOQ2NPO3TTUSXyYDjyLf

While Trump walked away with a $636 million profit from his crypto venture, nearly 1 million people who bought into it lost a total of $3.81 billion through the end of June. They were probably mostly his MAGA supporters, but to Trump they were suckers. https://trib.al/tQSlpdJ

The vibes at AIE world fair were ✨immaculate ✨

Come and chat with us in ICML 🥳 Excited to present Temporal Straightening for Latent Planning in the Tuesday morning session #1509. Let’s talk about world models, JEPA and representation learning.
agentic learning ai lab: Agentic Learning AI Lab @agentic_ai_lab presents at #ICML2026 in Seoul✈️🇰🇷 Come check out our latest research!
📍Main Conference Papers
Temporal Straightening for Latent Planning
Tue Jul 7 10:30am Hall A #1509
Context Tuning for In-Context Optimization
Wed Jul 8 5pm Hall A