AI in manufacturing enthusiast. CAD/CAM techie and weekend woodworker. 🤖🪵
Liverpool
Joined May 2023
- Tweets6K
- Following312
- Followers194
- Likes11.9K
Do you see teams actually switching models mid-project, e.g. one model for schematic capture and another for DFM checks? Would be interesting to know.
Come test drive all these models in @Flux!
We’re big believers in giving our customers access to frontier intelligence and the freedom to switch between models without having to switch their entire agent environment.
Different models have different strengths, and those strengths change constantly
Our job is to give them the best possible hardware engineering environment, regardless of which model is behind the wheel.
Will Hughes retweeted
Everyone assumes AI-written code at scale means a giant pile of slop. Me too at times.
Axiom just put out the best counterexample I've seen.
They took the Stacks Project, the reference every algebraic geometer works from, built up by the math community over two decades, and turned the whole thing into Lean.
Millions of lines, and every single one is checked by a computer. It either holds up or it doesn't compile.
What gets me isn't even the size, it's the structure. Stacks is this massive web where every result leans on a bunch of others.
Axiom compares it to the glass pyramid at the Louvre, all these individual pieces that only make sense once you see how they fit together.
They kept all of that intact, just in a form a machine can check.
They ran their agents like a software team to do it, with some owning the architecture, some the core definitions, some the individual proofs, and humans at the top.
To be clear, the math is still human work.
People wrote Stacks.
What's new is that it's now a verified foundation AI agents can actually build new math on.
Link is below ↩️
🤝 Paid partnership
Will Hughes retweeted
those of you who dont want to watch 19,000 shill videos 😂 can be assured I'm just finishing up a video about PPA.
Make sure you're subbed. It's a good one.
youtube.com/@lostintech3d?su…
Will Hughes retweeted
Opus and Sonnet are right now giving **lots** of usage in $20/$200 sub plans, but... at API costs (which enterprises pay), GPT models absolute own the pareto frontier.
Both in terms of time taken and cost.
The latest (Luna+Sol+Astra) combo is far more efficient than Claude 5.5 family.
Will Hughes retweeted
🤖 How can companies turn AI investments into real business results? We’re bringing you another great conversation from @Gartner_inc! @ScottwLuton sits down with Sanjiv Gupta of @Aptean to discuss turning AI ambition into measurable business value.
Catch the latest interview and subscribe to the Supply Chain Now YouTube channel for more conversations from Gartner: hubs.ly/Q04D3cQ80
#SupplyChainNow #GartnerSC #Aptean #AI #SupplyChainPlanning
Will Hughes retweeted
When slowing down makes you faster...(& smoother...)
ADB (Autonomous Decentralized Bearing) is a patented technology developed by Sosuke Kawashima (Coo Space Co., Ltd.)
By introducing a small indent into the outer race of a deep groove ball bearing, the balls become self distributing & non-contacting without the need of a cage.
Translation: the balls don't bump into one another so no cage is required, which then means sliding friction is virtually eliminated, energy consumption is reduced by over 90%, and no lubrication is necessary.
That precise indentation I mentioned slightly decelerates and then accelerates each ball back to full speed as it passes over it. A small gap is thus created between the balls due to this change in speed. Any incidental contact between the balls acts as a restorative force, effectively transferring momentum between them and ensuring uniform spacing naturally.
Pretty cool stuff (and so very simple) in other words.
This technology is currently touted as being applied in deep sea & competition skate/bike/wheelchair cases, but seems like a prime candidate for any super low friction, demanding environment, and/or long life applications.
(All original video footage from Sosuke Kawashima on yt)
People want the AI products but not the power plants behind them. 244 GW of the pipeline now sits in states that don't want it either. Where permitting takes years, capacity just goes elsewhere. No surprise.
Every deal announcement: 'category leader', '$550B TAM', '$1B synergies'. I care less about the deck and more about DSP plus optics under one roof. Co-packaged optics only really works that way.
$385 for a Pi? At that point just buy a mini PC and be done.
Raspberry Pi hikes prices again, with the Raspberry Pi 5 16GB hitting an eye-watering $385. hackster.io/news/raspberry-p…
Will Hughes retweeted
A bit late to the party, but if you’ve been eyeing a Bambu Lab printer, their Black Friday deals definitely are worth a look.
If you’d like to support the content, shopping through my affiliate link earns me a commission at no extra cost to you: modbotarmy.com/go/bambubf
Will Hughes retweeted
Every Linux server with TCP listeners can be exploited to root level remote code execution.
Working proof of concept shown by Anthropic researchers.
Wondering why 9.8 and 10.0 actually.
Damn this one is bad.
Will Hughes retweeted
i tried reaching out to autodesk but couldn’t get anywhere. i would love to record an episode with them discussing this and where they think the future is going
Will Hughes retweeted
Is that a ghost?
No, it's "ghost imaging" incorporated into an X-ray approach!
Scientists created a computational imaging method that reconstructs X-ray pictures from a tiny fraction of the photons normally required in #medical diagnostics: bit.ly/4BA6Z8T
#Halloween😉
![An image acquired using the new method (a) is compared with conventional radiology (b) under identical photon counts. [Image: Tiqiao Xiao, Shanghai Advanced Research Institute, Chinese Academy of Sciences]](https://pbs.twimg.com/media/HV-G1NAXwAABJRp.jpg?name=orig)
ALT An image acquired using the new method (a) is compared with conventional radiology (b) under identical photon counts. [Image: Tiqiao Xiao, Shanghai Advanced Research Institute, Chinese Academy of Sciences]
I'd take the tooling and tribal knowledge within 300 miles of Detroit over another SaaS round any day. The talent is there, the capital isn't. ngmi.
California will never concede that it is overdone and full of itself, and we cannot wait for it to do so.
But as long as Midwest capital is still entranced and existing mfg aloof to startups, the solo-preneur grind / involuntary stealth is all that's left
start-midwest.com/news/the-s…
Will Hughes retweeted
i don’t think it’s the use of AI as the cause, but i had to switch away from spotify because it’s become unusably buggy over the last year
Will Hughes retweeted
I open-sourced 43 agent skills for shipping real products. No coding.
Everything I learned from 190K installs and sign ups.
Why was it needed?
Popular AI engineering skill collections are for people who work with the code a lot. They skip discovery, go-to-market, and product analytics.
I wanted to create something for people focused on the user and business value. A set of skills that helps you:
→ research the market and competitors
→ listen to the users, e.g., on Reddit
→ identify your value proposition and value curve
→ decide how to grow
→ learn from people using your product for real
Then agents build, review each other's work, and ship. You approve the release.
Inside, 6 packs from the first idea to paying users:
→ Product (8): discovering what to build
→ AI shipping (10): a reliable, secure AI-built app you can sign off on
→ App stack (14): sign-in, payments, email, analytics, and legal pages
→ Agent teams (4): other agents as partners, e.g., Codex
→ Agent harness (6): a reliable setup, with you in the loop
→ Work humanizer (1): work writing that doesn't read as AI
You don't have to memorize them. Type /start-here and it explains where you are, what's done, and the next steps.
Free repo (GitHub): github.com/phuryn/product-en…
If it helps, star the repo so the next person finds it.
Apollo put humans on the moon. This buys GPU clusters that depreciate in 3 years. And I'm sure it'll all pay off.
.
Big tech’s AI spending dwarfs America’s biggest historical mega-projects
In just two years (2025–2026), the four largest tech companies are planning to spend more on AI infrastructure than the United States spent on the Manhattan Project, the Marshall Plan, and the Apollo Program, combined.
All figures are adjusted to 2025 dollars
Manhattan Project (1942–1946): $36 billion
Marshall Plan (1948–1952): $137 billion
Apollo Program (1960–1973): $189 billion
Total of the three historical projects: $362 billion
Now compare that to AI capital expenditure from the four hyperscalers:
Amazon: $332 billion
Microsoft: $308 billion
Alphabet (Google): $276 billion
Meta: $207 billion
Total AI capex (2025–2026): $1.123 trillion
That is more than three times the combined cost of those three historic American mega-projects.
Source: @eco3min
Will Hughes retweeted
Run AI workloads at the edge with significantly lower power consumption.
Analog compute-in-memory (aCIM) technology is changing how AI inference is performed by processing data directly within memory, reducing the energy and data movement associated with conventional architectures.
Our subsidiary Silicon Storage Technology (SST) developed memBrain SAGE IP using embedded SuperFlash® technology to enable energy-efficient AI processing in ASIC and SoC designs.
Discover how AnalogAI is leveraging this technology to develop ultra-low-power edge AI processors designed for on-device inference and adaptation, with applications in robotics, autonomous vehicles, industrial automation and more.
Read how analog AI and compute-in-memory are advancing energy-efficient edge intelligence: mchp.us/4stKAWB.
@EEJournalTFM #EdgeAI #AnalogAI #ComputeInMemory #Semiconductors
