akiselev

The weights are in SRAM, so the LLM architecture is burned in but the weights can be updated.

The tokens per second performance numbers coming from Cerebrus/Talas are several orders of magnitude higher than models running on GPUs, which is such a huge step change that it will enable many more uses of LLMs that are impractical otherwise. I.e. think about gamers and burning in an LLM chip on a game console like a future Play Station - it doesn't matter if its a frontier LLM if it allows them to talk to in game characters in real time and have the LLM drive the storyline. The devs may be limited to training/RLHFing legacy models, but the performance enables many more use cases.

Knob the Yob has a certain catch to it.

I think at that point the phrase "engineered soil" loses all utility. We've been engineering soil with domesticated herd animals since prehistory, bringing fertilizer from pasture to arable land at the very least. If we look further at the most recent archaeological research on cultivation, there's growing evidence that soil engineering is how societies move from cultivation-assisted hunter gatherers to fully sedentary agriculture (and the strongest evidence, i.e. from extant isolated tribes in jungles, is that even the so called hunter gatherers participate in extensive soil engineering to support cultivation).

That's my fault, I should have prefaced that I'm just talking about the US. I have no idea what the situation is like in Europe (for some reason I assumed biofuels weren't big there). Due to US density and geography, most marginal land here wouldn't be returned to little more than pasture. It depends on the state but most of that land was never forest to begin with.

Most crops in the modern world run an engineered soil anyway.

What do you mean by engineered? The most fertile places in the US (i.e. the southwest) run on multi-million year old alluvial plains where micronutrients are deposited from mountain runoff. NPK and some micronutrients are supplemented but the most fertile regions tend to be the least "engineered". The engineering goes into the massive irrigation projects, not the soil, precisely because engineering the latter is so much harder.

The vast majority is grown on marginal land, just above pasture. They can't grow better crops without massive works of engineering and tons more fertilizer and energy use. The alternative is to just use slightly less of that land, because the animals are going to have to replace that feed from somewhere. Distillers grains are valuable because the fat and protein are used for finishing cattle for human consumption in feedlots so the sugars are either going to the cows or the biofuels.

The "limited value" isn't so limited when we're talking about an additive to gasoline. The first thing we tried polluted the entire world with a background level of lead!

So it's wasteful, not at all "Green", and inefficient, but do we have a replacement octane booster that wont poison people?

I'm not sure it's all that wasteful. The waste product from biofuel production is distillers grains [1] which are just fed back to animals afterward for the protein, fiber, and fat content.

[1] https://en.wikipedia.org/wiki/Distillers_grains

From the article:

DD6 is a second-generation nickel-based single-crystal superalloy developed by the institute with fully independent intellectual property. Its chief engineer, Li Jiarong, said the alloy’s performance matches or exceeds that of comparable second-generation superalloys used in Europe and the United States, at a lower production cost.

US manufacturers have already developed sixth-generation SC superalloys and most Western airlines are on engines with third- and fourth-generation materials.

The technology behind single crystal superalloys is relatively well understood, the problem is getting the process reliable enough to be economical in an industry that requires tens if not hundreds of billions of dollars to develop through trial and error. The TFA's point is that unlike EVs or semiconductors, the turbofan industry is between a rock and a hard place that China's other successful industries weren't.

When I ask it to tweak recipes and stuff, it frequently says stuff like "my favorite way to..." or "I really like [x]".

I have a viscerally negative reaction to a machine claiming it has a favorite anything.

I remember working with the Pegasus at Caltech in the early 2010s for a tech demo that was never going to launch.

I'm amazed it's still operational.

That's a very vague prediction that took decades to bear fruit. The concrete predictions behind the investments into companies like Pets.com and Webvan failed. It took the survivors like Ebay/Paypal and Amazon to build the digital payment and shipping infrastructure over decades until cultural acceptance hit critical mass.

Because proteins famously aren't molecules.

Tertiary and quaternary protein structures are much more complex than molecules and have emergent properties.

I only use it to help hold together some complex assemblies when putting them together so I can't speak to it's strengths, but:

Do you need a conductive substrate on both sides to apply the release voltage?

Yes, you attach VCC to the substrate you want to remain bonded and GND to the substrate you want to detach.

How available is it for purchase if you aren't Apple or Samsung?

You can just call up your usual 3M distributor and request it. It's over $2,000 for a 100 meter roll so it's not something you'll usually find in stock at your local Grainger but it's not some super-secret material only available to the biggest manufacturers.

The vast majority of the time, you don't hear about it at all. Leveraged buyouts and the monopolistic strategies like buying out all the private doctors or veterinarians in a region get all the negative press but they're a tiny fraction of private equity.

The big money is in really boring industries like mining/oil/resource extraction, power plants, infrastructure, construction, and other industries that are predictable and in high demand everywhere. PE firms often get the best deals because they thrive on those kinds of connections and can offer up large amounts of capital on favorable terms in exchange for first dibs. The "rich get richer" is their primary strategy and it works without minmaxing exploitation because that's a bottom feeder strategy, not one that can guarantee steady returns on tens of billions of dollars.

CPU cache space for code is much smaller than GPU memory for models (and the former is more important for performance since many CPU operations like pipeline parallelism are latency bound, not compute bound).

Based on requirements.txt it uses build123d so OpenCascade is the geometric kernel (CAD engine backend)

The real power with these kinds of tools isn’t prompting one shotted models but giving agents the ability to do the full workflow. You give them a description of the part and how it’s supposed to mate with parts from McMaster, Misumi, existing parts libraries, etc and the agent downloads the models, asks any clarifying questions to clear up ambiguities (using available part configurations to provide options when applicable), uses measurement tools to validate the design, provide material details for FEA, read and use PDF drawings/datasheets, and so on.

At least, that’s the theory. The problem is that none of the existing CAD tools (almost all exclusively built on Parasolid) are set up to support agentic workflows. None have proper text based representations, with the possible exception of OnShape’s feature script which is too undocumented and proprietary to be of much use. Even if it was supported, Parasolid isn’t set up to provide the kind of detailed error reporting needed to provide agent feedback.

I’ve been experimenting with this in ECAD by giving agents the ability to edit Altium files directly and it’s been working very well (even with footprint drawings!), but my attempts to do it with MCAD have fallen flat on their face because it’d require developing a geometric kernel from scratch with this workflow in mind.

Curious to hear if people have use cases where they find 1M works much better!

Reverse engineering [1]. When decompiling a bunch of code and tracing functionality, it's really easy to fill up the context window with irrelevant noise and compaction generally causes it to lose the plot entirely and have to start almost from scratch.

(Side note, are there any OpenAI programs to get free tokens/Max to test this kind of stuff?)

[1] https://github.com/akiselev/ghidra-cli

They can oneshot relatively simple parsers/encoders/decoders with a proper spec, but it’s a completely different ballgame when you’re trying to parse a very domain knowledge heavy file format (like the format electronics CAD) with decades of backwards compatible cruft spread among hundreds of megabytes of decompiled Delphi and C# dlls (millions of lines).

The low level parts (OLE container, streams and blocks) are easy but the domain specific stuff like deserializing to typed structs is much harder.

Totally agreed. I’ve been reverse engineering Altium’s file format to enable agents to vibe-engineer electronics and though I’m on my third from scratch rewrite in as many weeks, each iteration improves significantly in quality as the previous version helps me to explore the problem space and instruct the agent on how to do red/green development [1]. Each iteration is tens of thousands of lines of code which would have been impossible to write so fast before so it’s been quite a change in perspective, treating so much code as throw away experimentation.

I’m using a combination of 100s of megabytes of Ghidra decompiled delphi DLLs and millions of lines of decompiled C# code to do this reverse engineering. I can’t imagine even trying such a large project for LLMs so while a good implementation is still taking a lot of time, it’s definitely a lot cheaper than before.

[1] I saw your red/green TDD article/book chapter and I don’t think you go far enough. Since we have agents, you can generalize red/green development to a lot of things that would be impractical to implement in tests. For example I have agents analyze binary diffs of the file format to figure out where my implementation is incorrect without being bogged down by irrelevant details like the order or encoding of parameters. This guides the agent loop instead of tests.

I used the latest submissions from sites like crackmes.ones which were days or weeks old to guard against that.

When I was developing my ghidra-cli tool for LLMs to use, I was using crackmes as tests and it had no problem getting through obfuscation as long as it was prompted about it. In practice when reverse engineering real software it can sometimes spin in circles for a while until it finally notices that it's dealing with obfuscated code, but as long as you update your CLAUDE.md/whatever with its findings, it generally moves smoothly from then on.

There’s not much difference, really. I stupidly didn’t bother looking at prior art when I started reverse engineering and the ghidra-cli was born (along with several others like ilspy-cli and debugger-cli)

That said, it should be easier to use as a human to follow along with the agent and Claude Code seems to have an easier time with discovery rather than stuffing all the tool definitions into the context.

The book Brett uses as his main source, Waging A Good War, is an incredible book that I strongly recommend. It treats the Civil Rights movement as a military campaign and analyzes it from the perspective of a military historian.

Not in the sense that it was viewed as a war by the protestors, but in the sense that the logistics, training, and operations of the Civil Rights movement were a well oiled machine that looked like a well organized, but nonviolent, army (including counterexamples where there was no organization).

One of the most memorable details is how James Lawson trained in nonviolence under Ghandi and came over to train protestors in nonviolent tactics. They gathered in church basements to scream insults and spit on each other to prepare for the restaurant sitins and other ops.

Shameless plug: https://github.com/akiselev/ghidra-cli

I’ve been using Ghidra to reverse engineer Altium’s file format (at least the Delphi parts) and it’s insane how effective it is. Models are not quite good enough to write an entire parser from scratch but before LLMs I would have never even attempted the reverse engineering.

I definitely would not depend on it for security audits but the latest models are more than good enough to reverse engineer file formats.

We can finally get rid of all that middle work. That adapting layer of garbage we blindly accepted during these years. A huge amount of frameworks and libraries and tooling that has completely polluted software engineering, especially in web, mobile and desktop development. Layers upon layers of abstractions that abstract nothing meaningful, that solve problems we shouldn’t have had in the first place, that create ten new problems for every one they claim to fix.

I disagree. At least for a little while until models improve to truly superhuman reasoning*, frameworks and libraries providing abstractions are more valuable than ever. The risk/reward for custom work vs library has just changed in unforeseen ways that are orthogonal to time and effort spent.

Not only do LLMs make customization of forks and the resulting maintenance a lot easier, but the abstractions are now the most valuable place for humans to work because it creates a solid foundation for LLMs to build on. By building abstractions that we validate as engineers, we’re encoding human in the loop input without the end-developer having to constantly hand hold the agent.

What we need now is better abstractions for building verification/test suites and linting so that agents can start to automatically self improve their harness. Skills/MCP/tools in general have had the highest impact short of model improvements and there’s so much more work to be done there.

* whether this requires full AGI or not, I don’t know.

I think it’s an Adobe Flex app now.

The real question is whether “debugging” the LLM is going to be as effective as debugging the code.

IME it pays dividends but it can be really painful. I’ve run into a situation multiple times where I’m using Claude Code to write something, then a week later while working it’ll come up with something like “Oh wait! Half the binaries are in .Net and not Delphi, I can just decompile them with ilspy”, effectively showing the way to a better rewrite that works better with fewer bugs that gets done in a few hours because I’ve got more experience from the v1. Either way it’s tens of thousands of lines of code that I could never have completed myself in that amount of time (which, given problems of motivation, means “at all”).

I'm self taught but used to work as an EE professionally doing high speed digital and RF mixed signal work.

It honestly might have been easier without that experience because KiCad is open source and their S-expr file format is easy to use. I'm stuck with Altium since that's what I learned on and am used to.

Most common question on HN seems to be "what are people building". Well, stuff like this.

Hear, hear! I’ve got my altium-cli repo open source in Github as well, which is a vibe coded CLI for editing vibe reverse engineered Altium PCB projects. It’s not yet ready for primetime (I’m finishing up the file format reverse engineering this weekend) and the code quality is probably something twelve year old me would have been embarrassed by, but I can already use it and Claude/Gemini to automate a lot of the tedious parts of PCB design like part selection and footprints. I’m almost to the point where Claude Code can use it for the entire EE workflow from part selection to firmware, minus the PCB routing which I still do by hand.

I just ain’t wasting time blogging about it so unless someone stumbles onto it randomly by lurking on HN, they won’t know that Claude Code can now work on PCBs.