Hottest possible takes for electrical engineering:
decoupling capacitors mostly don’t matter
RF is actually way easier than chuds say
ferrite beads don’t do anything but cover up terrible design
you should never use 2 layer boards
Supply chain is 100x harder than design
ταῦ retweeted
We all know that to get the best performance in FPGA IO timing you need to pack the register to the IOB.
This week I am looking at why we do this, which is obvious and why sometimes we cannot which is less obvious and how we can correct it should we fail to get the register to pack due to how we have implemented our design.
adiuvoengineering.com/post/m…
Replying to @edelopo
Thank you, I knew this needed more discussion, but I didn't want to make the post too long, so thanks for giving me the chance to clarify:
1) There are no needs, only wants. Nothing is infinitely valuable; every choice is a trade-off among the things you value, including (but not limited to) getting home from the airport, the other uses of the money you would spend on the ride, and everything else you could do with your time and resources.
When the price of an Uber rises, the people who still take one are those for whom the ride is worth more, relative to their other wants, than it is to the people who decide not to take one. That is what “who needs it most” means in any operational sense. The claim that someone who will not pay the higher price “needs” the ride more is simply an assertion that your ranking of their wants should override the ranking revealed by their own choices.
2) Who has more money is not arbitrary. In a market economy, income is largely the return on production that other people value enough to pay for. The person who can more readily spend $200 on an Uber is creating more value for other people than the person who cannot. Giving the ride to the higher bidder therefore tends to allocate the scarce resource in the way that best satisfies society's needs/wants (call it what you want).
3) Almost everyone at the airport can find $200 if the alternative is truly intolerable. The fact that many choose not to pay it does not show that they lack the money; it shows that, given everything else they could do with those dollars, they do not value the ride at $200. Their choice reveals the intensity of their want relative to their other wants. Price rationing simply respects that ranking.
------
@"Item 3) is illegal in most places (you need a license to serve as a taxi)"
True, and that IS the problem. The market should be much freer in that respect.
ταῦ retweeted
Every time I think European voters cannot be more in denial about pensions, they prove me wrong: they are even more in denial than in my worst nightmare.
The average French person believes there are 6.4 workers for every retiree. The real number is about 1.8.
ταῦ retweeted
Wrote a new blog post on building a fully custom open source networking switch ASIC chip.
talesonthewire.com/projects/…
ταῦ retweeted
I designed a radio chip with Claude in 7 days. Alone it would’ve taken months.
But how do you use AI without skipping the learning?
On Friday I’m asking Harald Pretl and Carsten Wulff, analog designers and professors experimenting with AI.
What should I ask them?
ταῦ retweeted
TSMCのAI Design Kit(ADK)とは何か? AIエージェントによる半導体設計は新しい段階へ - Vengineerの妄想 vengineer.hatenablog.com/ent…
ταῦ retweeted
"Embedded System Design" by Peter Marwedel
An awesome, open access book
link.springer.com/book/10.10…
ταῦ retweeted
Tohoku University and Toyota-supplier Aisin built a 16 nm AI chip (SoC) with 32 MB of embedded STT-MRAM that holds both a compact Linux boot image and Tiny YOLOv3 (object detection AI model) weights on-chip.
No external DRAM or SSD is needed for startup or inference. It boots in 65.7 milliseconds (39× faster) and runs object detection at ~50.7 mJ/frame with big energy savings. They used weight sharing and hybrid 8/16-bit math to fit the model.
The chip also includes on-chip SRAM (used for feature maps and other working data) and a DDR interface, so external memory can still be attached when larger models or more capacity are needed. TSMC manufactures the chip.
Nice real-silicon demo of near-memory computing for edge AI — useful for automotive cameras, sensor fusion, and always-ready devices that need instant-on and low standby power. Not revolutionary: Others – Hailo, Ambarella, Renesas, etc. – already use large on-chip SRAM, embedded MRAM, or in-package DRAM, but it’s a good example of the gains when OS + model live in non-volatile on-chip memory. $TSM $AMBA #Hailo #Renesas #semiconductors
ταῦ retweeted
Hydrogen-1.0 gives SOTA performance on FrontierScience-Research and on par to SOTA on GPQA-Diamond at an API cost similar to GLM-5.3. On our internal research scenarios, Hydrogen-1.0 scores higher than or on par with GPT-6-Astra at maximum thinking on scientific computing (76.9% vs 71.8%), forward reaction (90.5% vs 85.7%), and NMR structure elucidation (26.7% vs 30.0%).
read the full blog at blog.deepprinciple.com/intro…
ταῦ retweeted
There are a number of computational primitives one can implement with oscillator-based computers, e.g. solving combinatorial optimization problems, and associative memory.
Higher dimensional oscillators can also implement the attention, which forms the basis for modern LLMs
ταῦ retweeted
Computer Chips Are the Most Magnificent Things Humanity Has Ever Made
newyorker.com/magazine/2026/…
ταῦ retweeted
80 IQ: chips are beautiful and beautiful things have value
100 IQ: claudeslop about what layer the value will accrue in the AI era
120 IQ: chips are beautiful and beautiful things have value
Computer Chips Are the Most Magnificent Things Humanity Has Ever Made
newyorker.com/magazine/2026/…
ταῦ retweeted
I used Divyansh's approach to build an entire Nintendo into ~1GB of Qwen3.5 weights.
Qwenputer, meet Nesformer: github.com/dbreunig/nesforme…







































