synapsomorphy

Thanks, fixed

Full stack / Generalist Engineer

  Location: US (GA)
  Remote: Yes
  Willing to relocate: Only to PNW or Bay Area
  Technologies: Python, Linux, embedded C, SQL, React, ML (PyTorch, LLMs), electrical

Hi! Generalist here looking to work on hard and meaningful problems with cool people.

2 YoE, coming off a year building the entire tech stack for a YC alum deeptech + biotech startup, from scratch, as the sole SWE or EE. (product: https://andsonbiotech.com/platform/)

I've also built/worked on rockets, drones, a lunar lander, a Mars helicopter at NASA, a LLM-driven car, AI benchmarks, and new type of 3D printer. Read about these: https://synapsomorphy.com/

Prefer early-mid stage, say seed to C, and core products that have 1 or more of [ML, biology, hardware], in order of preference.

Full stack / Generalist Engineer

  Location: US (GA)
  Remote: Yes
  Willing to relocate: Only to PNW or Bay Area
  Technologies: Python, Linux, embedded C, SQL, React, ML (PyTorch, LLMs), electrical

Hi! Generalist here looking to work on hard and meaningful problems with cool people.

2 YoE, coming off a year building the entire tech stack for a YC alum deeptech + biotech startup, from scratch, as the sole SWE or EE. (product: https://andsonbiotech.com/platform/)

I've also built/worked on rockets, drones, a lunar lander, a Mars helicopter at NASA, a LLM-driven car, AI benchmarks, and new type of 3D printer. Read about these: https://synapsomorphy.com/

Prefer early-mid stage, say seed to C, and core products that have 1 or more of [ML, biology, hardware], in order of preference.

Full stack / Generalist Engineer

  Location: US (GA)
  Remote: Yes
  Willing to relocate: Only to PNW or Bay Area
  Technologies: Python, Linux, embedded C, SQL, React, ML (PyTorch, LLMs), electrical

Hi! Generalist here looking to work on hard and meaningful problems with cool people.

2 YoE, coming off a year building the entire tech stack for a YC alum deeptech + biotech startup, from scratch, as the sole SWE or EE. (product: https://andsonbiotech.com/platform/)

I've also built/worked on rockets, drones, a lunar lander, a Mars helicopter at NASA, a LLM-driven car, AI benchmarks, and new type of 3D printer. Read about these: https://synapsomorphy.com/

Prefer early-mid stage, say seed to C, and core products that have 1 or more of [ML, biology, hardware], in order of preference.

Maybe the quality is reduced, sure. But if you "do some searches" you can find all of those things for any major software release.

Seems to me like people in Apple's walls are forgetting that the outside world is not some Garden of Eden. But yeah, I'd have to use it to say for sure.

I'm a Linux and Windows user thinking of getting a Macbook, mostly for the hardware.

All these recent proclamations of disappointment in Tahoe seem insanely overblown to me. The problem that this post leads with is that thumbnails' corners are too rounded, which "misrepresents" the original? Seriously?

Maybe it's worse now compared to the golden years, I don't know, never owned a Mac. And it's fair to criticize it from that perspective. But I am completely at a loss for how any of these issues could be bad enough to make you switch platforms. Windows and Linux are not exactly usability all-stars! I had to write my own app for decent speech-to-text on Linux which is built in at a system level on Macs.

This feels to me like just the age-old tale of people wanting to (love | hate) brands, when really, things are nuanced. I switched from Android to iOS recently and the experience did not change much. iOS is absolutely not "borderline unusable" like I've seen many claim. If anything it's maybe a 10% nicer experience overall.

Lack of nuance in people's takes makes for less signal in the noise and makes it annoying to figure out the actual pros and cons of different platforms.

It's an arms race between human writers and AI. Writers want to sound less like AI and AI wants to sound more like writers, so no indicator is reliable for long. Today typos indicate a real writer, so tomorrow LLMs will inject them where appropriate. Yesterday em dashes indicated LLM, so now LLMs use them less.

Beyond these surface level tells though, anyone who's read a lot of both AI-unassisted human writing as well as AI output should be able to pick up on the large amount of subtler cues that are present partly because they're harder to describe (so it's harder to RLHF LLMs in the human direction).

But even today when it's not too hard to sniff out AI writing, it's quite scary to me how bad many (most?) people's chatbot detection senses are, as indicated by this article. Thinking that human writing is LLM is a false positive which is bad but not catastrophic, but the opposite seems much worse. The long term social impact, being "post-truth", seems poised to be what people have been raving / warning about for years w.r.t other tech like the internet.

Today feels like the equivalent of WW1 for information warfare, society has been caught with its pants down by the speed of innovation.

Assuming Eric / Core doesn't come out with some scathing "real story":

Well, it's better to figure this out today (that Eric / Core are not so great) rather than a year or two down the line when I'd have already bought a new Pebble. Still sucks, I was excited. Never had one but I want something in the same niche.

Does anyone have suggestions for other good low-capability, long battery, hackable eink watches?

Chinese builders are not equal to Chinese hackers (even if the hackers are state sponsored). I doubt most companies would be interested in developing hacking tools. Hackers use the best tools available at their disposal, Claude is better than Deepseek. Hacking-tuned LLMs seems like a thing that might pop up in the future, but it takes a lot of resources. Why bother if you can just tell Claude it's doing legitimate work?

Survived 102 seconds before a rug pull. Impressive!

That issue is the fourth most-reacted issue, and third most open issue. And the two things above it are feature requests. It seems like you should at the very least have someone pop in to say "working on it" if that's what you're doing, instead of letting it sit there for 4 months?

I'm thinking a lot about the ARC-AGI ML benchmarks, especially the "shape" of the dataset and what that says about how it should be solved. I think there's good reasons to believe that deep learning - at least differentiable SGD backprop style - is a bad fit for this specific benchmark, due to the tasks being almost entirely discrete symmetries, and also having so little data to approximate the discrete symmetries with continuous ones (considering deep learning to be the learning of continuous symmetries). I think that a more explicit and discrete approach is the way to go, and it's possible to build something surprisingly general and not heuristic-based even without gradient descent, guided by minimum description length to search for both grid representations and solver functions. I'm looking for teammates for ARC-3 so hit me up if this sounds interesting, I'd love to chat!

I made a viewer on my website to build intuition for my preferred perception algorithm which is entropy filtering + correlation. Pretty neat to check out the heatmaps for random tasks, there is a lot of information inherent in the heatmap about the structure of the task: https://synapsomorphy.com/arc/

Interesting essay. But it attributes something magical (ability to solve undefined problems) to humans and says that AI doesn't have it.

It seems pretty meaningless and not engaging with the real problem to say that AI doesn't "actually" write movie scripts or paint pictures. Like this doesn't line up with my definitions for doing those things which AI clearly fulfills.

And human intelligence arises from a well defined problem: maximizing f(environment, self) -> babies.

Also: if it were possible to measure, which it isn't, I strongly suspect that ability to solve well-defined problems and ability to solve poorly-defined problems are highly correlated, not totally uncorrelated. Happiness is a poorly defined problem, but it's just one of many, and has its own pile of things to consider that can isolate it from the general ability to solve poorly-defined problems.

I do like the framing. seems to be describing something similar to Goodhart's Law.

I'm completely puzzled on why space-based compute is so exciting to everyone all of a sudden. I have worked on spacecraft and the constant power benefit seems comically far from outweighing the many, many negatives, even if launch cost is zero, which we are still very far from.

Am I missing something? Feels like an extremely strong indicator that we're in some level of AI bubble because it just doesn't make any sense at all.

Full stack / Generalist Engineer

Location: US (GA)

Remote: Yes

Willing to relocate: Yes, only to PNW or Bay Area

Technologies: Python, Linux, embedded C, SQL, React, AI/ML (PyTorch, LLMs), electrical

Résumé/CV: https://synapsomorphy.com/resume.pdf

Email: On CV

---

Fullstack engineer with 2 yrs exp. Coming off a year building the entire tech stack for a YC alum deeptech + biotech startup, from scratch, as the sole SWE or EE. (product: https://andsonbiotech.com/platform/)

I've also built/worked on rockets, drones, robots, a lunar lander, a Mars helicopter at NASA, and a new type of 3D printer. See portfolio on website: https://synapsomorphy.com/

Prefer early-mid stage, say Series C or earlier, and core products that have 1 or more of [biology, ML, hardware], in order of preference. But mostly I just care about solving hard and meaningful problems with cool people.

Not sarcasm at all.

There are some "AI thinkers" who are trying to make 8% faster CUDA kernels for attention, and there are some trying to save the world. These fields are called "capabilities" and "alignment". There is some overlap but not much.

LessWrong is mostly the latter, and labs and universities are mostly the former. That said, many LessWrongers work on AI at labs or universities or other places.

LessWrong.com - this is where virtually all of the serious AI thinkers are.

As an extreme believer in AI (to the point of being a doomer and expecting+fearing ASI) I have to agree with this wholeheartedly.

We have to be careful to avoid blaming AI as a technology for the incredibly hamfisted way it’s being implemented in most products, and affecting online spaces.

Maybe there will be a bubble burst. That doesn’t mean AI won’t eventually transform the world.

It’s hard to believe in nuance but also very important.

Kilobytes or single digit megabytes. It happens because Sharepoint sporadically alters created/edited metadata for any (?) file it stores. Most programs don't care about that but Solidworks does.

Sharepoint is one of the worst, most bug-ridden softwares I've worked with.

It has a bug with Solidworks (3D design suite) that sporadically makes files completely un-openable unless you go in and change some metadata. They are aware of this, doesn't seem to be any limitation preventing them from fixing it, and it has sat unfixed for years.

Microsoft's cloud storage as a whole is an insane tangle where you never know where you'll find something you're looking for or whether it will work. Some things work only in browser, some only in the app, zero enumeration of these things anywhere.

Completely unsurprised and I'm sure there are many more vulnerabilities ripe for the picking.

This is a really interesting discovery. In ants it's apparently common for one species to stop being able to produce workers on their own, and use the sperm from another species instead.

In this case, that happened. But if you do that, you can only expand as far as the other species expands. So you can expand further if you can find a way to keep the males of that species around with you.

This species does that by having a reproductive pathway that, if a queen is fertilized by that 'domesticated' species, the DNA of the 'host' species is removed from the eggs. So you get an ant that has none of the host's DNA. Except they do inherit the mitochondrial DNA (it always comes from the mother). The 'domesticated' males and the 'wild-type' males do look slightly different - it's not clear if this is because of the mitochondrial DNA or because they're raised differently or what.

I read someone compare the domesticated species to a 'superorganism organelle' - just like an archaea cell sucked up a bacteria to become a eukaryote, the host species sucked up the domesticated species to become some combination of both.

Wild to think what other crazy ways of living and makin babies must be out there that we haven't figured out yet.

The accuracy is going to be the real make or break for this. In a paper from 2018 they reported 92% word accuracy [1]. That's a lifetime ago for ML but they were also using five facial electrodes where now it looks confined to around the ears. If the accuracy was great today they would report it. In actual use I can see even 99% being pretty annoying and 95% being almost unusable (for people who can speak normally).

[1] https://www.media.mit.edu/publications/alterego-IUI/

I'm honestly kind of surprised there haven't been significant large-scale attempts to well-poison LLMs with certain viewpoints/beliefs/whatever. Maybe we just haven't caught them.

Entirely the second. When Vera Rubin starts reporting its regular scans this will be made very clear because we'll probably find 10+ interstellar objects per year at minimum.

It appears to me like the linked explanation is also subtly wrong, in a different way:

“This is why a flat surface like a sail is able to cause lift – here the distance on each side is the same but it is slightly curved when it is rigged and so it acts as an aerofoil. In other words, it’s the curvature that creates lift, not the distance.”

But like you say flat plates can generate lift at positive AoA, no curvature (camber) required. Can you confirm this is correct? Kinda going crazy because I'd very much expect a Cambridge aerodynamicist to get this 100% right.

o3's cost was sliced by 80% a month or so ago and is also cheaper than Claude (the output is even cheaper than GPT-5). It seems more cost efficient but not by much.

Nit: Diffusion isn't in place of transformers, it's in place of autoregression. Prior diffusion LLMs like Mercury [1] still use a transformer, but there's no causal masking, so the entire input is processed all at once and the output generation is obviously different. I very strongly suspect this is also using a transformer.

[1] https://www.inceptionlabs.ai/introducing-mercury

I fabbed a couple FreeEEG32 boards [1] recently and have half of a design for my own board put together.

This technology CANNOT effectively move a mouse around on a screen today, much less control robots. If it could, they wouldn't have to implant things in paralyzed patients' brains just for basic computer control.

I do think it's a very interesting field and there's a lot of improvements to be made. It's also extremely sensitive to noise (for best signal you can't be anywhere near mains power), any movement of facial muscles completely drowns out the brain signal, and getting electrodes prepped properly is time consuming and requires skill. And even in optimal conditions the SNR is not amazing.

I'm looking into dry, active electrodes as well as inter-electrode impedance detection to solve electrode prep, and a driven right leg circuit to help SNR. TI ADS1299 (ADC used by this and most other hobbyist EEG boards - directly targeted at EEG thus fairly expensive) has impedance detection and a DRL circuit but best I can tell neither is used by most boards [2]. I'm also interested in pogo pin electrode arrays for increasing spatial resolution.

Honestly most EEG boards seem, to me, more for show and money than anything else. No one even attempts to quantify noise levels, and they have very large margins for basically being breakout boards for ADC chips. (and $300+ for a fabric cap with passive electrodes???) And no one who says "look at all this stuff you can control with EEG!" has any projects of actually controlling anything with EEG, because it's extremely difficult. They just link to old papers where someone put together a control system slightly better than random chance.

Would love to collaborate on something here if anyone has any interesting ideas, I think hobbyist EEG could be done a whole lot better.

[1] https://github.com/neuroidss/FreeEEG32-beta

[2] This one does implement impedance measurement which is nice.

Location: Atlanta, GA

Remote: Yes

Willing to relocate: Yes

Technologies: Python, embedded, PCBs, Linux, Torch, Postgres, TypeScript, React, CAD, 3D printing

Résumé/CV: https://synapsomorphy.com/resume.pdf (website: https://synapsomorphy.com/)

Email: patrickwspencer at gmail dot com

I spent the last year building the entire stack for a hardware deeptech recent YC grad - backend, frontend, PCBs, firmware, infra, etc, as a one man dev team. I'm looking for somewhere with experienced hackers to learn from and somewhere I can wear a lot of hats. Prefer hardware (especially robotics - just built myself a robot arm to dive into it) but open to any hard + fun problems!

Atopile is another thing in the circuits-as-code space: https://github.com/atopile/atopile

As a half EE/half SWE I think there are significant benefits to circuits as code but I'm not impressed with this one. Atopile has a narrower focus (autorouters are really really hard) and doesn't use as many buzzwords. Like why on earth does a "web first approach" matter at all for hardware development?

But also, GUI tools are getting better, Kicad 9 had a lot of changes that made templating / reusing blocks easier. And it works fine if not great with version control.

I don't see circuit-as-code taking off with humans anytime soon, it's much better but not enough better to convince EEs many of which don't code much or at all. But I can see it becoming much more common as LLMs get better at complex circuits.