Attempt #1

For a first attempt, I crushed a few ripe berries (A) in 3:1 ethanol:vinegar, spotted it out on a TLC plate, and ran it with 2:4:4 acetone:ethanol:ethyl acetate. Ditto for partially ripe berries (B), unripe (C) and leaves (D). The first run was way overloaded, and even repeats with smaller drops (just A, B, and C on a single plate to be thrifty) were a bit of a mess. You can see red-fluorescing chlorophyl, bright blue (phenolics, flavinols etc) some lots of other junk. The purple anthocyanins from the ripe berries were also visible, and various things stained yellow after a dip in some KMnO4 (with NaOH and K2CO3) stain.

The problem is that we’ve extracted all sorts of things - we need a way of excluding as much of this junk as possible so that we can see the glycoalkaloids we’re interested in. Plus, I think the TLC mobile phase needs to be more polar…

Take 2 (and 3 and 4…)

With some suprisingly helpful advice from Fable 5.1 (hooray for more nuanced bio filters!) I tried again with a more refined extraction protocol:

• Mash up the berries in some vinegar (~4% acetic acid)
• Spin down and push the supernatant through a syringe filter to get it nice and clear
• Bring the pH up to ~11. The vinegar buffers it for a while and then it shoots up, I overshot a few times. What’s cool is that you get a color change to watch this - yellow appears then fades as you add the NaOH until suddenly it stays. Fun doing titration at ~1ml scale haha.
• Cool then spin down, leaving a little pellet at the bottom
• Wash with water, spin down
• Dissolve in ethanol (I tried plain and lightly acidified)

The theory here is that we’re extracting the alkaloids in acid then crashing them out at high pH in their (mostly insoluble) free-base form, then washing away most of the salts, sugars etc that remain soluble before re-dissolving the target molecules in ethanol. Some proteins and things will crash out too at the high pH, but this way we (hopefully) have waaaay less junk to deal with.

I then ran spots of these extracts (switching to A=ripe, B=unripe and C=Bittersweet Nightshade, a different species that isn’t edible and should have lots of the target compounds) with various mobile phases, e.g. 4:1 ethanol:water, or variants thereof with some ethyl acetate added (and sometimes vinegar) - e.g. 5:3:2:1 EA:ethanol:water:vinegar.

Besides tweaking the mobile phase and extraction, I also experimented with different ways to get the compounds to show up nicely - since they aren’t flourescent, staining or charring were our bet bets. There is Dragendorff’s reagent, which stains alkaloids specifically, but I don’t have it. For the charring, I added some hydroponic pH DOWN (phosphoric acid plus a bunch of buffer salts and such) to ethanol. The salts crashed out. I’d dip the TLC plate in the phosphoric acid+ethanol then place it on a hotplate set to 180C and wait for spots to go brown.

This was one of the more frustrating parts - I’d be able to see faint smudges by eye, but filming them on the pure white background of the plate was extremely challenging! I ended up relying on image manipulation to make the spotches stand out. Worse, the KMnO4 stain fades fast and changes in real time! Still, eventually, with charring or staining or both, I was able to get a few variants with obvious blobs in B and C and fainter smudges in A - possible candidates.

Lessons Learned

• Nice to learn a little more about acid-base extraction
• I got much better at doing consistent spots on the TLC plates (a micropipette set to 3ul helped a lot)
• I learned that alkaloids get ‘smeary’ in TLC if they’re not fully protonated, or fully de-protonated. So, one usually adds a little ammonia to the mobile phase to keep things basic. I didn’t have, but made some with NaOH+NH4NO3 in a vial such that the ammonia vapor hung around in the TLC chamber; mixed results. OR, you can just acidify things, hence the vinegar in some runs :)
• I learned from my X thread discussion that you can just taste solanine, bro - and indeed, with some very careful tasting and the bittersweet nightshade as reference, I think I can pick it out, and not taste it in the ripe fruit
• I learned the ripe fruit are pretty tasty (I only ate one) and some are already pretty well bred for low solanine, hopefully one day we get a nice domesticated variety that is extra safe and tasty to much.
• I got to practice my (very rusty) chemistry, mixing up stock solutions and adjusting pH and making yellow chemistry E&F references and somehow spilling NONE of the stain or nasty lye or anything!
• Fable 5.1 (which came out that day) was suprisingly helpful - I tested it on this expecting instant bio refusal (we are extracting a toxin after all) but instead it was a great teacher.

Mixed results

The first image in this post was the one from the day that I thought had the best chance of being what I wanted: some blobs in B and C that weren’t in A. But they were streaky and indistinct.

As I wrote this up, I figured I’d give this one more go: more concentrated spots, a fresh batch of phosphoric acid+eth with more acid, EA+eth+vinegar as the mobile phase. This got much better brown spots, further up. And, prompted by codex, I tried something with this and yesterday’s ones that I hadn’t before: UV after the acid+heat stage - and got some glow where there wasn’t before! So maybe those higher spots in some of the runs were the compounds after all? Except from the lit I expect them to have low Rf?

Anyway, I end this post as I ended the video: inconclusively. If we wanted to do this properly, we’d get some proper stain, and some ammonia, and some green potatoes as another reference, and run everything a little more carefully… but I’m about ready to be done with these berries for now - after all, this ‘quick test’ has already taken more than a day!

I hope you found the journey interesting :) LMK what mistakes I’ve made, I bet there are plenty.

until next time,

Johno

PS: Title may be misleading, apparently this (taxonomically complex and messy) group might have more solasonine and solamargine vs solanine, so “hunting for steroidal glycoalkaloids in nightshades” might be better given ambiguity about exact species/accession and target molecules haha.

Let The Gradients Flow: Bilinear Splatting and Sampling

When you have a particle with some x,y position, say, 7.2, 8.5, one way you could ‘draw’ it or have it deposit ‘pheremones’ to the fixed grid of the substrate is to round off and use the nearest pixel. BUT this throws away some signal - how are you supposed to get a gradient that would nudge the particle a little to the left? Instead, it’s generally a good idea to distribute the signal around the nearest pixels based on distance:

The same goes for sampling. Each particle reads some info from its neighborhood - by doing so with this bilinear sampling approach, we can get signals like “it would be better if that signal was stronger to the left” (hand waving here obviously).

Perception, Communication, State

We have a bunch of agents (say, 4096) and a shared grid (say, 96px square). Each agent has a position, velocity, a color it can deposit, and some internal state. They ‘perceive’ their neighborhood, then choose how to modify their directions and the pheremones + colors they are depositing.

Importantly, the agents can’t see eachother directly (this makes life a lot easier - training boids or something where each particle reacts to other nearby particles requires a LOT of tricks to make things performant). Instead, they can interact indirectly, by reading pheremones from the grid. I also ‘splat’ a density signal to the grid, which they can read to see how many other agents are/were nearby.

For perception, traditional slime sims use only two or three ‘antennae’ - to give mine a better chance of learning, I expanded to a 5x5 pixel grid that they can see. You can have this aliogned to the agent’s direction of movement, or “world aligned” (I tested both) - I kept it world aligned for ease of computation.

I also gave them a few gradients, one radial, one vertical and one horizontal. This feels a little cheaty, but not implausibly so. Otherwise we could end up with rotationally invariant setups that would really struggle to produce a target image. The jax code gives a nice concise look at the key bits:

class State(NamedTuple):
    field: jax.Array       # [H,W,communication]
    canvas: jax.Array      # [H,W,RGB]
    pos: jax.Array         # [agents,xy]
    heading: jax.Array     # [agents,xy]
    hidden: jax.Array      # [agents,memory]
    pigment: jax.Array     # [agents,RGB]

def perceive(state):
    density = density_field(state.pos, CFG["grid"], CFG["grid"])
    expected = CFG["agents"] / CFG["grid"]**2
    density = jnp.log1p(density / (expected + 1e-6))
    world = jnp.concatenate([state.field, state.canvas, density], axis=-1)
    sample_pos = state.pos[:, None, :] + OFFSETS[None, :, :]
    local_patch = bilinear_sample(world, sample_pos.reshape(-1, 2))
    local_patch = local_patch.reshape(CFG["agents"], CFG["patch"]**2, -1)
    cue_center = bilinear_sample(CUES, state.pos)
    global_inputs = jnp.concatenate([state.heading, 2 * state.pigment - 1, cue_center], axis=-1)
    return local_patch, global_inputs


def brain(params, local_patch, global_inputs, hidden):
    spatial = jnp.tanh(jnp.einsum("nkc,ks->ncs", local_patch, params["spatial"]))
    encoded = jnp.concatenate([spatial.reshape(CFG["agents"], -1), global_inputs], axis=-1)
    encoded = jax.nn.silu(linear(params["encoder1"], encoded))
    encoded = jnp.tanh(linear(params["encoder2"], encoded))
    hidden = gru(params["gru"], encoded, hidden)
    trunk = layer_norm(hidden, params["ln_scale"], params["ln_bias"])
    trunk = jax.nn.silu(linear(params["trunk1"], trunk))
    trunk = jax.nn.silu(linear(params["trunk2"], trunk))
    out = linear(params["head"], trunk)
    turn = CFG["max_turn"] * jnp.tanh(out[:, 0])
    speed = CFG["base_step"] * jnp.exp(0.35 * jnp.tanh(out[:, 1]))
    emission = jnp.tanh(out[:, 2:2 + CFG["comm"]])
    gate = jax.nn.sigmoid(out[:, 2 + CFG["comm"]])
    proposal = jax.nn.sigmoid(out[:, 3 + CFG["comm"]:])
    return turn, speed, emission, gate, proposal, hidden

Learning Something Interesting

I tried re-creating some image textures using a texture loss as in the fantastic uNCA work (todo link). This works as long as the texture you want is blobs :D After poking at it for a while I was almost ready to call it on neural physarium, but then the day before my call with Ehsan to show (lack of) progress on our silly quest, I gave it one more go with codex, brain dumpimg all my ideas and challenging it to get some shapes growing. Success!

Then, with further pushing, we were able to get persistent painting working too:

The agents deposit pigment, which gradually fades and diffuses, so they must learn to keep it topped up, too. It is even somewhat robust to peterbations too, since we train with some degradation. CHeck out this tweet to see a video of me playing with an interactive toy I made to show this.

I use the pooled training as developed in the original NCA distil.pub article. The mushroom emoji is the best of the lot - thanks to the constant motion, finer details or more complex shapes can be a bit of a mess. Most of my exporiments looked more like this:

Still, better than nothing!

An Evolving Way Of Doing Research

I didn’t really write code for this project. Sure, I seeded it with some older code of mine, and described things in pretty fine detail as we went along. But Codex did the hard work - and this meant things got done that I would never have patience for. Lots of tweaking of the loss function to eke out extra gains. Easy ablations of things like rotating the perception vs keeping it world aligned. HTML research logs with lots of nice animations, interactive demos, and loss curves. Parallel training runs to test things using model for compute. I felt more like a thesis supervisor with an amazingly productive student, rather than a researcher in the trenches. And, as A.M. quipped when I said I’d made them a jax port: frameworks don’t really matter these days! Wild times.

Anyway, I’ll try to make a video on these too, but for now I hope this blog post at least serves as a minimal sort of write-up/reference to point at. This is a fun and unexpored area, I hope this inspires you to give it a poke too. Good luck, LMK if you train anything cool - J

Why Now?

• AI is getting v. good
• Software skills may have a limited remaining shelf life
• I want to step back and see if I’m in the best spot to contribute vs doing work that will happen anyway
• It was good for me to have a break from ‘participation’, to sit back and relax and spectate rather than being plugged-in to each model release and AI news story. I think I was pretty burnt out, as many in the space are

Factors To Consider

• Do I enjoy it (e.g. I like teaching)
• Does it have lasting value (e.g. teaching some specific tech << teaching some general principle)
• Does it have +ve impact (I want to do good)
• Are other people doing it? (I prefer less populous areas of the map)
• Does it pay? (I’ve historically skewed unprofitable-but-fun, now worry about financial footing given how many of the skills I have might tank in value as they get automated)

With that, let’s look at a few options.

1) Back To The Mines (AI Research)

Lots of effort and capital is flowing into AI research right now. Many things feel ‘overdetermined’ - that is, they’re going to happen pretty much regardless of what I or any other individual chooses to work on. That said, even if you think AI research is on track to be fully automated in the next few years, you can’t look at the current state of things and not see that there is still LOTS of work to do:

I’m a little rusty on model training, but I have > a decade of experience at this point, and am sure I could find ways to contribute. There are organizations working to build AI that benefits people, doing hard + interesting technical work, that pay decently. I’m not sure the day-to-day of begging codex to fix distributed training code would be my favourite thing to do, especially if I don’t have a compelling reason to want the result NOW (vs a little later, without me having to lift a finger). Still, I’m chatting to a few people who may convince me that their org is doing cool enough stuff that I’m compelled to chip in and help it along :)

2) Frontier?

• Probably not, unless alignment or application like bio.

3) Bio?

• Slow and expensive
• Possibility (e.g. ML 2014) to bring it to many more people
• …and build on models getting better, cheaper synthesis etc in the future
• got some ideas…
• but not quite yet. Can’t afford to pursue myself, but can keep on side (or beg for funding as a startup)

4) ‘Poke The World With Science’ course

• Didn’t want to do a course
• But came up with an idea I would be OK charging for, that I love
• Participants get a kit of science stuff each [time period] and some lessons on a few things, then self-directed exploration using those principles to take stuff further and build up a tech tree for themselves over time
• more on this soon possibly

Shorter-Term Projects

• Hardware Hacking Evals (time sensitive, impactful, I can’t afford to try it on my own dime)
• Portland DIYBio scene (exciting, I think we could build a commuinity here in time)
• Electronics Kickstarter (might test the waters as a way back to EE)

Why Not Answer.AI

I had a great time at Answer, they’re good people! Solveit was fun, although (personal opinion only) it feels like we made a fantastic tool for coding with AI in 2024, or learning to code with AI in 2025… As AI progress relentlessly continues, I feel the need to make some different bets. I’m sure they’ll continue to have fun teaching AI to code like Jeremy! And it’s good that there are people trying different things as we all explore what the heck programming looks like now and in the future :)

You Tell Me!

There are lots of areas I’m probably unaware of or underestimating! I’d love to hear from you if you have ideas for what I should work on! Even just my announcement on X that I’m formally unemployed has led to lots of lovely chats from people wanting to share the cool things they’re working on, which I am loving. I really appreciate everyoine who has taken time to talk to me and answer my unusual questions as I’ve been setting out to figure this stuff out.

Also, more directly than ideas, you (dear hypothetical reader) can steer what I do with the application of money. If there’s a direction you think I should go, a project you think I should do, and

Final Musings & FAQ

One of the biggest sticking points when I talk through my decision space with people is my lack of motivation to work on things I think are going to get done. Someone is going to build it, why shouldn’t it be me? There are two main pieces where I think I am unusual as regards motivation: - The puzzle is important: the fun part for me is figuring out if something can be done, and how. Once I can see that I lose a lot of drive - and for many software problems now the ‘how’ is some variant of “wait for the AI to get good enough then ask nicely”. This takes a lot of the wind out of my sails. Obviously it is still worth doing the things, when you care about the outcome or will get paid to do it or whatever… but I care above average about having to figure out new stuff. - I am anti-trend. I really dislike working on things that lots of other people are working on for some reason. (This is a bad preference for career earnings!). I am so much happier if I can find something that wouldn’t otherwise happen, open up some new area of the map and then make it easier for others to come check it out and explore with me. This is why I loved AI art in 2020 and lost interest as it took off, why I’m drawn to niche things like neural cellular automata or DIY biology… IOW my explore vs exploit dial is set way over on ‘explore’.

Another frequent question is, given my penchant for teaching, why not teach courses? Once I find something cool worth teaching, I am allergic to monetizing it. I can’t stand charging for content - possibly because I grew up in a poor country where a price tag meant no access? I’ve also been very resistant to monetizing my YouTube/teaching content for fear of losing control over what I cover. If I have a paycheck attached to what I produce, that will heavily incentivize pushing towards the content in demand, chasing the hype. I might set up a Patreon this week and try to overcome these instincts to some extent, but this is why the list abive doesn’t feature things like running courses on agentmaxxing or whatever the current thing is.

Money Talk

People don’t like discussion salaries for some reason. Here are some figures to help you understand how I think about money:

• I grew up in Zimbabwe where many people live on less than $1 a day, so everything in tech sounds insane.
• My first work was stats+ML+coding for low hundreds of dollars a month (for a good cause)
• Out of college I was doing data science for low hundreds of dollars a day (for a good cause)
• My salary at Answer.AI doing AI R&D was something like $150,000 a year.
• (This split e.g. last year to living: ~$50k, tax: ~40k giving: ~25k, IVF & savings the rest - very approximate)
• I’d consider short consulting for $500/hr if it’s a few hours, possibly less if it’s a longer time. Also, people who know me know that I can be expoited to work for free if you can nerd snipe me, feel free to exploit this information.
• I’d been giving myself $500/month as a lab budget upper bound, that might have to drop while we’re drawing down savings. SO, if you want to contribute to me doing science stuff specifically, donating for lab spending specifically would make a huge difference to what I can do - this is also one reason I am thinking of starting a patreon even though it would be unlikely to move the needle much re: more general living expenses. Guilt-free budget for science is the dream :)

Given the pace of improvement in AI coding, I am more inclined now than ever before to let money become a factor in what I do with my time. Right now, I am valuable for getting things to happen with computers, likely to be accelerated further as tools improve. But within a few years AI may well make my involvement less necessary - so those skills at least (i.e. software engineering specifically) feel like they have a couple of years left for me to capitalize on them.

The Hardware

The ‘motion platform’ is an Ender 3 V3 SE 3D printer that has served me well over the past few years. A Raspberry Pi hosts the site and sends G-CODE to the printer. The micropipette sits in a 3D printed holder with a Dynamixel XL430 servo and control board in charge of pushing the plunger. You can swap in a lower-volume micropipette for more precise work. The servo is one I happened to have on hand from a past robotics project, a little pricier than necessary for this application I suspect. Paper is held to the bed by magnets. original conception

The Code

To turn this into a demo I waved a magic wand~ asked Codex to whip up an interface based on the previous code I used to give the bot a control API. The resulting code is here but comes with no warranty, not even the implied warranty of ‘a human has read this at some point’. My opinion on interfaces like these is that it’s most fun to roll your own to your exact specifications - the latest coding models speak web UI and g-code well enough to bash something together in minutes. This demo was completed from a voice prompt + a couple of follow up requests while I rode the bus to the Portland Art Museum on Wednesday :)

Operation

You draw a picture, which gets added to the queue. It is ‘painted’ by dispensing droplets of liquid in a grid pattern - by default they sort of sit there on the paper. You can have fun by 1) wetting the paper first using the spray bottle of water for some wet-on-wet watercolor vibes, 2) spraying afterwards 3) dabbing with paper towel / tissue to get rid of excess liquid or 4) whatever you like.

If I’m not there, odds are the printer is in semi-auto mode. Put a piece of paper in the marked spot, stick it down with magnets on the corners, then hit the central scroll knob/button thingee on the printer interface and it should start on whatever the next drawing in the queue is. Given the crowd, I expect some surprises and take no responsibility for what happens.

Other Things We Might Have Spoken About

The recent microfluidics tests

The purple tomato seeds are from a bioengineered tomato from norfolk healthy produce, are not for sale but can be shared with your local community (a lovely approach): https://www.norfolkhealthyproduce.com/pages/faqs

The duckweed has been transformed using agrobacterium carrying this RUBY plasmid: https://atinygreencell.com/products/pcambia2300-rubygreen

The DIY gene gun works but is waiting for full tests, here is an early WIP post from a few months ago.

The DVD hacking adventure has a writeup here

The e. coli engineering project has a writeup + video here

Anything else is probably linked somewhere on https://johnowhitaker.dev/ or https://x.com/johnowhitaker. My gmail is also johnowhitaker if you want to get in touch that way.

JULY update:

Tween 80 is the opposite of Span 80 (the surfactant I tried to order) - I assumed they were synonyms! 5% Span 80 in mineral oil and suddenly I could easily do water-in-oil droplet generation :) I have also successfully tested laying down plastic on PCBs rather than glass slides and bonding on the cover slip - this opens up the possibility of easy holes + electrodes, I’ll be ordering some boards to try these ideas shortly. Droplet pics here: https://ln.johnowhitaker.com/entry/61

First Contact:

Inside the enclosure is a slimline SATA DVD drive module (model DS-8ABSH, which you can also buy standalone on Amazon for ~$30). It is connected to a USB-to-SATA adapter that has also been crammed into the enclosure, with a USB cable that you can then plug into your laptop. It shows up as a disk drive, and speaks SCSI - a standard way computers communicate with storage media of various kinds. It’s not uncommon for manufacturers to use extra, secret, undocumented SCSI commands as a sort of hidden communication channel for things like firmware updates - and so initial work included messing about seeing if there were any that produced interesting responses (and ejecting the drive or causing weird noises a lot in the process).

At some point we dug up an old firmware update program, designed for DOS and Windows. Poking around in the binary, along with our scans, revealed a few special SCSI commands - including one which gave some Lite-On specific metadata, and one that let us dump a 1MiB firmware window. The firmware was encrypted, but the EXTRAINQ info gave the pieces needed to derive the AES key and IV and decrypt it, giving us our first look at (some of) the code running in the device.

The first chunk of that was 8051 assembly code, which I could load into Ghidra to decompile and feed to Codex to start exploring. We quickly spotted some promising-looking bits - now the question was, could we modify this and start running our own code instead?

Updaters in The Matrix

We needed to know the full sequence of steps involved in a firmware update, and after some guesswork based on static analysis of the DOS updater got us nowhere, we decided to run the program in a little simulated world where we could inspect exactly what it was sending during the process.

Pretty soon, we had the process mapped out well enough to replay an update with the firmware we’d dumped. But when we tried to make modifications, nothing happened. The process would run fine, we could even read the modified firmware back out - but nothing we did seemed to change much. In desperation, we even thought that the USB-SATA bridge might be filtering some key command, and I opened up an old MacBook from a friend to wire in the drive directly - no luck.

Alas, it wasn’t a missing ‘commit’ command after all: it turns out that there’s a firmware integrity step that happens before the drive allows new firmware to be written, and we were failing it. I even had a pretty good guess as to which set of bytes was the key checksum/stamp of approval, but it wasn’t anything simple that we could guess at. Here’s one place where having a tireless AI agent came in clutch. Codex searched relentlessly, finding a handful of updaters for different variants of this drive and others in the same family. Then it patched and re-patched our wine shim, adding more and more functionality as needed to convince these update utilities that they were programming drives, extracting out 5 or 6 ‘sibling’ firmware images (each with their own encoding nuances and tricks required) so that at the end we had a set of similar binary files we could compare.

Sure enough, those 14 bytes seemed like some sort of key/checksum, but no theory we could come up with revealed what, and despite a ton of work, we couldn’t find anything in any of the updater programs to help us out, nor could we see any code that checked it in the visible 8051 firmware. Still, at least these sibling images would come in handy as references later. For now, we faced a block - if we can’t modify the code, the project is sunk!

Papers Please: Bypassing Firmware Integrity Checks

We discovered that the official updater sends a separate profile-tail payload, which is essentially an 8051 helper overlay. By changing three bytes in this helper, we turn off a conditional and force the success path - sidestepping the firmware validity check entirely! So, now we can write whatever firmware we like…

Code Caves and Trampolines

The next step requires us to tread carefully. We’re planning to modify firmware that we don’t understand - a wrong step could leave the drive unresponsive and ‘bricked’. The standard approach here uses two delightfully named concepts. A code cave is a region that we can safely mutate - for example, we have a little 221-byte chunk of FFs at 0x6ee3 that can be safely over-written. We can write a piece of code there, and then patch some function somewhere to jump there (hence ‘trampoline’), execute our code, and then bounce back to where it was.

For example, a cautious first step added a delay function in the code cave and then triggered it during the update process while timing the different stages, giving a measurable timing difference. Next was reading out a single bit by conditionally setting an error state. Finally, we made something actually useful - patching the currentboot identity handler (0x4fc9) and using our code to modify the identity string. Now, we could issue a SCSI request and get back a customized response, which let us dump out data such as memory snapshots much faster than the one-bit-per-firmware-update early attempts.

Later on, this same idea was used to test all kinds of different hypotheses - we can do anything - as long as it fits in 221 bytes :)

Hardware Twiddling and Mysterious CDDs

We can see some 8051 code, but there’s also this big chunk that is encoded in some weird way. We tried a ton of different guesses, with no real luck. One thing stood out in the format/layout: a suspicious resemblance to some DVD encodings. The theory is that it uses some internal silicon decoder as part of the process.

It took a while, but eventually we were able to map out some of the code path responsible for ‘decoding’ this into code/data at runtime, and build a helper to let us use the on-chip decoding path to slowly read out the materialized CDD contents one chunk at a time. Each run required power cycling the drive - fortunately, I’d long ago hooked up some circuitry to let Codex trigger a physical disconnect of the drive at will rather than relying on my meat fingers to do the boring labor. Unfortunately, those reads don’t see the full picture, only slices of the working window containing a few snippets of code. And our code runs in a different context to the normal execution, so couldn’t ever materialize the full, decoded block.

Still - the hardware cold reboot was useful for various parts of this. I later added another line that let me fake the switch that detects whether the sled is inserted, so that we could fake eject/insert operations with code as well. These hardware mods are useful for reverse engineering, but hopefully won’t be needed by anyone attempting to follow in my footsteps.

The State Of Things

TODO we still can’t fully decode CDDs. Can move sled around with standard SEEK(), and have hacks to spin drive, toggle laser and other party tricks but no proper full low-level control yet for things like focus / positioning.

TODO write better narrative arc, were we doomed, recovery, more details of various bits.

Actually, those TODOs capture it well. I have a few bricked drives, a few in various states of disassembly and functionality. And despite lots of head-banging, we don’t know how to fully decode or modify the controller code that would unlock further progress. I’m pausing this project for now.

Magic AI and The Inevitable Pit of Despair

~All of the code for this was written by Codex, with me steering and chiming in occasionally but otherwise mostly letting it go and making sure it had all the tools needed to explore ideas. This project specifically is one I’ve long had in mind as a sentinel of progress. Past models, even quite recently, made ~0 progress, so it was amazing to see codex (with some opus help) blast through the blockers one by one and make a ton of headway.

As with any AI project pushing up against the ceiling of what models are capable of, there came a time when progress began to slow and I found myself going in circles a little. It can be easy to get stuck here - “one more prompt will surely fix it” is a the natural feeling, one that I think causes lots of burnout and pain when people (myself included) miss the signs. I’ll be honest - this time around I noticed, but kept going past that point without snapping out of it - which is why I only have 2.5 drives out of four in working order!

When faced with this block, there are a few tricks you can try: - Wait for better models - Package up the core question and ‘ask the council’ - GPT 5.5 Pro, Opus 4.7, even Grok - you can ‘live in the future’ a little by leaning on the top tier models in aggregate. Sometimes this is just enough to break through some barrier - Actually do some work. A nice option if you do have the skills for it :) - Find a way around, approaching things from a different angle or working on some other aspect for a while.

In my case, a combination of fresh ideas from Pro, fresh motivation from me, and the addition of some DVDs to open up new code paths to explore broke the first deadlock and got us moving again, but now in the CDD decoding puzzle I sense a limit and will take a break rather than pushing in circles any further.

Conclusions

The original coastermelt project made a huge impression on me in high school - it’s a tour de force of RE, and introduced me to many ideas including the very concept of interactive disassembly, bit banging serial communications via a button/LED pins, the power of soldering on very tiny wires, etc etc. It inspired me to get into embedded engineering, and was one of the motivating factors behind my own RE journey - not that I ever reached anywhere close to scanlime’s level.

There’s a lot of hype around AI for cyber stuff at present, thanks to scary tales of Mythos’ exploits and high-profile vulnerabilities being reported at a scary rate. Seeing these models step through the RE workflows I’ve long associated with only a handful of experts was certainly eye-opening. We’re not at the “as good as an elite hardware hacker” level yet, but it suddenly feels in reach. And this heralds an exciting world where curious hackers can tinker with their devices, but also a scary one where less benign individuals could compromise the peripherals and embedded devices that surround us - and unlike your phone OS, it is unlikely that the good guys will be able to patch every old mouse, hard drive and wifi router sitting around in our houses!

Anyway, I hope you found this interesting. If you want to read more on how the code works, or try it yourself, I’ve dumped ~everything on github - start here. As mentioned, I’m hoping this is only the start and that at some point I’ll get proper control and make the hardware do some tricks… Hopefully see you in a future post soon :)

Additional Notes

I left this section as a TODO for a few months. At the time, I felt a bit let down that the project had ~ended without a clear win of an obvious visible green e.c. strain I could use for agar art. With the benefit of a few months break to give some perspective (I’m writing this now in July 2026) the project looks a lot nicer! I got to work through the stages of a genetic engineering workflow, re-writing little bits of the code of life! The end result isn’t perfect because, shocker, biology is complex… but still :) I’m very glad I got the chance to try this, and who knows, maybe at some point I’ll give it another crack.

References:

• My blog (has a few posts with more info on this under ‘bio’): https://johnowhitaker.dev/all.html
• Frog paper: https://www.pnas.org/doi/10.1073/pnas…
• Various plasmid designs: https://github.com/johnowhitaker/BLY
• Extra paper on BBSs: “Structural and Functional Characterization of a Biliverdin-Binding Near-Infrared Fluorescent Protein From the Serpin Superfamily”

Errata: I’d written “ALA supplementation” down and said Alanine multiple times in the video - brain fart! Not a heme precursor. That would be Aminolevulinate.

Small Things

Here’s a few recent experiments and such:

I made an interactive site to explore tulip varieties (code, video demo). The UMAP clusters of the CLIP embeddings map nicely to color and locale, made it very easy to get a general vibe for the genus - I’ll definitely be copying this for other taxa!

My wife and I camped out by the Oregon coast and I took some photos of things under my UV flashlight. More pics on my iNaturalist once I take em off my SD card.

I extracted nicotine from tobacco and learned how to stain TLC plates with KMnO4, accidental science art :)

I tried an energy drink, crystalized some on a slide, used TLC to confirm that it does indeed have a ton of caffeine, and logged my vitals over the day.

I noticed that some bacteria have a lovely rainbow effect in the right lighting

Tried taking some caffeine out of coffee with activated charcoal, to some success! Image shows four references (1000, 500, 250 and 100 mg/L caffiene) and then some artificially strong coffee (left spot, rightmost plate) vs some that went through the charcoal decaf process (far right), with significantly less caffeine!

Others:

• Made a bad but fun theremin
• Filmed some fun glowy fluid dynamics
• Isolated a nice orange yeast https://ln.johnowhitaker.com/entry/27
• Grew out third and final (?) attempt at green e. coli, https://ln.johnowhitaker.com/entry/32, cloning strain, ran out of amp so transformation into gave BL21 questionable results, satellite colonies, no strong color yet
• Working on an explanation re: the green bacteria (WIP renders of the main proteins doing their thing) - video coming soon
• Agrobacterium transformation attempts on tomatoes, duckweed ongoing (another successful ish duckweed one worked for a couple of fronds, another attempt waiting for results)
• Tending to various plants, since it’s the start of spring here! Growing out some purple tomatoes, some berries, etc.

Sabbatical!!!!

Now the somewhat larger life update: I’m taking at least 6 months completely off work @ answer.ai. It feels like a particularly pivotal time in AI, and I want to reflect on what I should be working on going forward. Plus, I’d had an extended bout of sickness and was feeling a little burnt out and unproductive, so this will hopefully blow off the cobwebs.

So far, I’ve been really enjoying the wide open time (space) feeling. Not that work at answer was unpleasant - far from it - but ‘work’ is intertwined in my head with being productive, making progress; there are always outstanding TODOs. Having an empty agenda is a delightful contrast. As with all similar times in my life, my subconscious is frothing with a host if ideas that have been biding their time, possibilities that couldn’t quite justify themselves over the business. I’ve been trying to linger in this feeling, taking it easy, spending lots of down time walking and thinking and reading rather than rushing into the first ideas that pop up.

A less enjoyable note has been how quickly financial concerns popped up. We’ve got some runway of course, but now that my income has dropped to 0 I have to re-think a lab budget that was previously pegged at ‘a day’s income per month’ and figure out a balance between ‘not making the most of this time by trying to be too thrifty’ and ‘burning runway by spending too much’. I might have to factor in some notion of profitability into what I choose to do, something I’ve been fortunate at avoiding for most of my life. Such is the reality of living in a lovely but expensive place like Portland :) It’s not all downside though - I’ve already applied to one “micro-grant”, and making at least some work legible to such funding structures might impose at least a modicum of structure to an otherwise completely un-structured play time that could turn out helpful… But not yet, the focus for now is rest and recharge and fun.

As for directions I’m thinking of exploring, there are tugs in various directions:

• Everyone is running into AI applications, ‘agents’, etc. This strongly makes me want to run the other way - I much prefer exporing weird and ‘uncool’ spaces not the trendy thing.
• But AI is going to be hugely impactful, and I might still be able to steer it in ways I’d like it to go
• And I miss AI research. We’ve been very application-focused at answer, so I haven’t gotten to stretch my old deep learning + model training muscles in ages - there may be fun non-mainstream AI research work that could be fun, especially with coding agents there to skip some of the more painful parts of idea testing.
• (Maybe even some mechanistic interpretability work? More niche, still space for fun new ideas, relatively impactful?)
• On the other hand, I always love being in a field where I have a different set of skills to everyone else, so maybe I should get further from the AI core and instead help people without AI/coding/tech skills.
• And DIY Bio is fun and intellectually stimulating, also humbling+frustrating in its ability to not go quite as planned. I might see if I can get a chance to work with some bio people in a lab nearby if I can weasel my way in
• Making hardware used to be the thing I thought I’d end up doing. Now that I live in a country with McMaster Carr, Amazon Prime etc that might be a lot easier! Cons: does require some capital (as does bio) compared to messing with software, pros: holding real things in your hands and touching the world with your code is always cool.

I might have been a little more burnt out than I like to admit, closely watching the ever-advancing frontier of AI and trying to keep up + work on things that at times felt temporary given impending advances… Rather than committing to any of the above soon, the short-term plan is to dabble with everything, slowly - spinning plates, looking for things that are fun rather than projecting out to impact or income. This usually works out well :)

Anyway, it’s an exciting experiment. I feel really good about maximizing flexibility at this time, and look forward to seeing how it evolves. If you have ideas or want to chat, well, my calendar is as open as it’ll ever be, feel free to reach out! And if you want to watch how this goes, keep an eye on this blog and on X, I’ll try to keep logs of how this is going and what I’m up to. Until next time, J.

The Method

I take a small amount (10uL) of a slurry of diatomaceous earth (DE), add some (10uL, 50mM) calcium chloride and some (~1ug, in ~2uL) plasmid DNA and place it on a piece of parafilm stretched over a 3D-printed tube. I dry it in a filtered stream of air, then the tube adapts to the end of my cheap airsoft pistol, and I fire it at a leaf from a few cm distance.

The DNA I have codes for the ‘RUBY’ reporter, which expresses a red pigment. I extracted the DNA from bacteria carrying this plasmid which I got from Sebastian’s Biotech Bazaar. I used this miniprep kit to purify out the plasmid DNA.

Findings

I shot some tobacco leaves, some duckweed, a slice of carrot and some onion… In the leaves, around the spots shot with DNA (but not the controls) there are some, well, reddish bits that show up after a few days! And after lots of squinting at them I’ve convinced myself that at least some of these are cells expressing RUBY :)

My gallery is filled with blurry attempts to capture this, and none make it as obvious as I’d like. For example, there are also plenty of cases where, around an impact point where a clump of DE has hit, there’ll be some orange-ish crud welling up. Around most impacts (and all that I saw in the no-DNA controls) this stuff is pale white. Could this be RUBY too, expressed in some damaged cell(s) and being included in whatever is happening around the hole? Plus there are random red things just around the place - bits of fiber, random flakes - here’s a gallery of things I was more sceptical of:

The almost magenta dots in the carrot are apparently pretty much what RUBY looks like in carrot, that was a YOLO shot with some leftover stuff, I’ll do some follow-on ones with carrot callus that I have growing and hopefully get a clearer signal. But in general it was much harder looking for red/orange reporter in orange carrot tissue!

I didn’t see anything in the onion I shot. Nor did I see anything in onion that I tried my micro-needle idea on. But then, RUBY doesn’t work everywhere - for example this paper documents it being a poor marker in coconut (also a good ref for optimization strat once I can easily bombard and measure.)

For comparison: I’ve also been trying transformation with agrobacterium, which gives a much more macro result - here are some duckweed fronds expressing RUBY to various degrees:

Thoughts, What didn’t work, What Next

What didn’t work:

• Using graphite as the carrier (I wanted something more visible than the DE, so I sanded some graphite rod to get powder, but I didn’t ever see transformation using it)
• Using my tungsten micro-needles seemed to kill leaf cells, and when I tried them more gently on onion I didn’t see any transformation
• Bombarding duckweed was tricky (it tended to fly away in the blast) and I didn’t see any transformation of its tiny cells.

I think the red spots are enough to indicate that we’ve gone from 0 to 1. From here, we can optimise - playing with

• DNA, carrier and buffer amounts
• Nozzle diameter and geometry (I already switched to a converging-diverging nozzle which seems to give a much nicer shot). Ref for flow guiding barrel design (just found)
• Shot distance
• etc.

But I’m not happy with RUBY as a readout - too hard to see and measure. I’d love to use a flourescent reporter, but after months of trying I wasn’t able to get my hands on some. I have a construct with a fluorescent reporter ready to go for another project, which I’ll be ordering soon, for about $500… Once I get that I’m hoping we’ll be able to easily quantify how well a given shot does, and start optimizing everything. I wish individuals could order plasmids from addgene! Anyway.

Once you have some transformed cells, you can try to grow them out in tissue culture, optionally with something like herbicide resistance to select for the trasnformed ones. I’m hoping we can find a visual-selection-only flow since I don’t want to me making herbicide resistant plants or having other people work with them.

My dream is to work out a path for people who haven’t done this before, and don’t have the tools to do assembly and cloning, to still order a construct with their own protein/design, and get it into a plant, all as easily as possible. Wish me luck :)