Asher Baron Weintraub

(More) notes on AI

2026-05-13T00:00:00Z

I kind of grew up immersed in emerging technology. I was playing with 3D printers when they only existed in makerspaces and attended yearly conferences for DIY tech and virtual reality. At one point—as a twelve-year-old—I gave a talk to graduate students about that technology’s potential use cases across industries.

And I’ve always insisted on understanding the technologies I interface with (at least to some reasonable, greater-than-average degree).^[1] I played with AI as a developer around 2020, but lost interest quickly—maybe because it scared me.

I’m more of a cynic now, and when I published my last post on generative AI, I took the perspective of an outsider. It’s the first time I’ve watched the tech world’s fervor from the sidelines, and a new sensation for me. I wouldn’t have it.

Since then, I’ve made a concentrated effort to understand this world which had so rocked mine…

I’ve found that LLMs make effective coding partners, though non-technical users will find themselves trapped if they don’t understand the code they’re writing. I’ve “vibe coded” a couple of projects: an “assistant,” who emails prospective employers daily on my behalf; and an AI chat-like interface for ELIZA—but I’ve found LLMs most useful as a second set of eyes on human-written code, or for code-generation when given rigid project scaffolding.

But this is as much the creative’s game as it is the tech-savvy engineer’s. Even consumer tools show a capacity to democratize both technical and creative skillsets. But I fear a “taste singularity”—a point at which popular aesthetic preference stagnates, bolstered by a proliferation of AI-slop amalgam as the dominant form for creative output. It’s been comforting to watch the public pushback on AI’s use in creative fields.

For the time being, Industrial Generative AI usage lies primarily in the Technology sector, but the corporate world is betting on AI agents’ potential to fully replace middle management.^[2]

This AI ecosystem is predicated on the promise of “AGI,”^[3] or the assumption that language models will [soon] give technology the ability truly understand human meaning through linguistic morphology.

“Chat,” thus, is a poor interface for the medium. If computers understood us, why would we have them to write to us? Shouldn’t they just act? At a high level, interfaces shouldn’t be entirely text-based. It stopped being fun at Zork (long live Zork). From a technical perspective, this means that responses shouldn’t be so abstract. Large blocks of text are informative, but not especially conducive to multimodal usage. There’s a reason web APIs typically respond in a standardized language. We should be writing systems for highly adaptable consumption, across interfaces.

There’s a race to start companies which bring AI to existing sectors. Few of these will survive, though many will be “successful,” in getting acquired by larger companies. Tokenization is a bottleneck on the road to AGI,^[4] but the feverish pace at which infrastructure has been built means a large shift will need to take place. It won’t happen quickly.

I’m optimistic that AI can be used to better lives, alleviating the friction many of us face with the digital world. But the costs of AI, both monetary and ecological, are high.

You can imagine my joy when I learned that Bitcoin had 20x’d while I was piecing through blockchain infrastructure. ↩︎
Stanford 2026 AI Index Report ↩︎
Artificial General Intelligence ↩︎
Perhaps Byte-Latent Transformers are next? ↩︎

Notes on Generative AI

2026-03-06T15:30:00Z

As a creative, it’s hard for me not to find generative AI’s prevalence an attack on my humanity. I have long thought of art and creativity as an abstract human transformation of stimuli. Cutting the human out of this equation (or placing the human at either end of the process, rather than in it) is not inherently “bad,” but neither is it—by definition—creative.

It won’t take our jobs, but it will take our agency: we have a long track record of sacrificing anything and everything for convenience (read: each other, the environment, our health, our attention spans). Convenience, of course, is just the selling point to us laypeople—late capitalist bourgeoisie benefits financially by producing a new commodity. Using generative AI for creative means seems to me the next logical step in this process—the sacrifice of our humanity.

All the same, the ship has sailed on many of the sacrifices I mention above. If there is a way to use AI to reclaim some of this ground, I’m all for it. A facet of this of course, is reclaiming lost time—long since claimed by “the machine.” We can mindfully use generative AI as a productivity tool, preserving our collective conscience and without sacrificing our independent capacities for humanity.

Yamaguchi Framebuilding School: Day 11

2023-07-10T00:00:00Z

The last day! Mostly braze-ons and alignment; finishing (filing & sanding) is homework.

We started early at Ed’s request—8:30 instead of 9:00.

The workshop, when we arrived in the morning. Yamaguchi had set up our frames in vices, assumedly for no purpose other than a display of sorts.

This is another of Yamaguchi’s gravel bikes, which he showed us at lunch time. It’s got: an internally-geared hub (something I’m considering for my bike, though his has coaster brakes); a motorcycle-style weighted fork; a short wheelbase; and an engraving of the Yamaguchi crest.

After lunch, I almost decided to add down tube shifter mounts for the retro look I’m going for, but figured I could achieve the same with friction bar-end shifters, and avoid the ugly protrusions that would remain when not in use.

Cable bosses

My cable bosses. They all came with a ridge on the bottom, presumably for slotting into a frame with a matching design; I filed these off before sanding.

My frame in the vise. The brake boss angle doesn’t matter much, as long as they’re aligned. The decision is a combination of function—is the bike going to be carried on the rider’s shoulder—and aesthetics. I went for a slight angle off of the top for a clean, unobtrusive look.

Mr. Yamaguchi told me the cleanest look would be if the top of the bosses was perfectly colinear with the downtube & toptube. We positioned the frame such that the bosses were vertical at this point.

The distance between bosses doesn’t matter much, but we need enough clearance on either end to bend the housing into position cleanly. Yamaguchi’s suggested measurements are in the handbook he gives us.

Sanded, flux’d, and clamped on with the Cobra clips. It’s important to flux everywhere so that the surface of the boss doesn’t oxidize.

We did the same for the down tube boss…

…and the derailleur cable boss.

These were silver-brazed. Most of the braze-ons we did today didn’t need the extra strength provided by bronze; plus, finishing is easier—I’ll explain later on.

Disc tab

I was originally planning on putting a disc tab on the fork too, but it didn’t look great with the lugged dropouts and I figured it won’t really be necessary for the kind of riding I’m doing. Ed calls running a front disc and rear side-pull (or otherwise) “mullet brakes”–he called what I’m doing “reverse-mullet”. Fun stuff.

The disc tab didn’t quite clear the dropouts I’m using, so I spent a solid amount of time cutting and filing this shape. This is the end result, where it’s aligned in the jig (homemade, of course). There’s a minimum acceptable gap at all connection points.

Bronze-brazed. This section was tricky, since bronze (as I’ve learned) requires a pretty delicate amount of heat, but I wanted to melt new material without messing with the older dropout joints.

This is final or close-to-final. I went over and reheated the joints a few times to get rid of bumps and air pockets.

Mr. Yamaguchi tells me that there are dropout-disc tab combined lugs, but they don’t usually have rack mounting points like mine, and they’re often quite bulky.

Bottle cage mounts

Adding down- and seat-tube center lines using the square tool.

Once we marked the lower mount on both tubes, Mr. Yamaguchi held two bottle cages up to the frame where they would lie to adjust clearance. We made a new line slightly higher on the down tube to give a little more room.

We first use this punch tool to make an indentation and check positioning, then a stepped drill bit makes the hole.

The holes should be pretty perfect if they’re on the center line, but mine were a little bit off (I checked with a straight edge, held against the inserted mounts). I just filed at one side of each using a round file until they lined up.

Brazed! (Silver.)

Lowrider rack mounts

We put the fork in the vise, making sure it’s level.

Measuring the fork blades (tip-of-lug to tip-of-lug). 320mm.

Marking the halfway-point. I marked both sides of the ruler and connected the lines to account for the angle.

We then made a center line using the file.

…

Hole-punch method again.

Aaaaaand… brazed.

Repeat for the other blade.

Disc tab reinforcement bar

This miter cut is done the same way as the brake bridge, but doesn’t need to be exactly centered like the bridge. It’s silver-brazed on.

To my short-term but viceral dismay, I had gotten confused by the upside-down frame placement, and brazed on the rod upside-down. It doesn’t matter much, but the angle is shallower rather than steeper, and the slight taper in the tubing is misaligned with the diameters of the chain & seat stays. Oh well.

Polishing

Mr. Yamaguchi had me go at my frame with a metal wire brush drill bit. I couldn’t get the hang of it for a while, but it turned out I was just being too gentle—this was quite an aggressive task. It ends up getting rid of excess silver (since it’s softer than bronze), removing any discolouration from heat, and making the rest shinier. I’m sure I could have done a better job, had I spent the whole day on it, but the frame looks pretty good as-is, and it was already a bit more of a workout than I was hoping for.

Clamp slots

Cutting slots on the stem and seat tube for clamps is pretty straightforward; and I mean that literally: it’s two parallel hacksaw cuts, right in the center of the brazed-on clamps. My cuts were a little angled, so Yamaguchi took care of most of these for me. I see it as a future opportunity for some fun designs, like diamond/star-, arrow-, or heart-shaped slots.

Facing

We had to face the seat tube, head tube, and BB shell so that seatposts, headsets, and bottom brackets, respectively, will fit properly. It’s very easy for the tubing to deform throughout the framebuilding process due to heat, so this cannot be skipped.

Cutting oil, tapping fluid, or at least WD40 should be used when facing.

Seat tube

Facing the seat tube. This tool is called an “adjustable reamer” (seatposts are usually 27.2mm).

Head tube

This is just a headset installation tool, with a cutting attachment. It only goes in one way. We keep twisting until the larger teeth touch the top of the tube and make it shiny.

My head tube was slightly deformed at the bottom, so Mr. Yamaguchi put this old, conveniently-sized Campagnolo bottom bracket tool in to hold the tube’s shape, while he hit it with a hammer. This was equal parts funny and nerve-racking to watch.

Bottom bracket shell

My frame set up in the vise. When I asked why we haven’t used the wood blocks for cutting rather than c-channels, Mr. Yamaguchi responded, “so you don’t get spoiled.” He also told me they’re not accurate enough for cutting. Fair enough.

The old Campagnolo bottom bracket shell facing tool, in my frame.

The most important thing here, as with anything concerning the bottom bracket, is making sure the threading is going the right way. Since the shell is English-threaded, we rotate the tool towards the back of the bike (normal on the left, reverse-threaded on the right/drivetrain side).

It’s not hard for this tool to jump when forced, ruining the BB shell and requiring us to start over from scratch. Yamaguchi told me to exercise caution and, when it felt stuck, back it up part-way and then go forwards again—as many times as needed before it moved smoothly.

We did non-drivetrain side first, all the way, before backing it out partially and doing the same for the drivetrain side.

This is mostly a holder/positioning aid for the next tool, but it screws in much like a bottom bracket, using the tools on the counter behind it. Newer BB shell facing tools come with this step built in.

It works the same as the head tube facer, only it doesn’t face inside the shell. We just give it 10 or so revolutions—just enought to make the surface flat & shiny.

Shifting cable guide

We set up the frame upside-down, added perpendicular center lines on the BB shell, and then punched & drilled a small hole in the bottom of the shell.

We didn’t use the whole stepped bit this time—just the smaller portion, before switching to a different sized bit. I couldn’t get the drill to make a hole so Yamaguchi stepped in and did it for me. It seems like he operates the drill slower, but with more force, which was kind of the opposite of my approach.

We then tapped the hole to make a screw fit in. It’s very important to keep the tool vertical while doing this, so that the threads aren’t misaligned. The cable guide will mount here. Mr. Yamaguchi was ready to give me a plastic one, but I think I’m going to order an old steel Campagnolo one off eBay instead.

Alignment

This is kind of a funny step to me, since the measurements are so precise, but the only way to adjust alignment at this stage is wrestling with the frame. It’s important nonetheless.

Checking the fork dropouts are aligned. Spacing is checked & adjusted the same way as the rear dropouts, described below.

Checking alignment

Using the tubing height gauge, after adjusting the seat tube the same way we did a few days ago. Once aligned, it’s held in place with a special jack-like tool.

We check the rest of the tubing the same way.

Using a needle height gauge to make sure the head tube isn’t crooked. Even if the center is aligned, the top and bottom could be misaligned.

The tubing height gauge has this convenient protruding tube to make sure the dropouts are equally-spaced.

We used the digital calipers to check spacing.

Mr. Yamaguchi widening the dropouts. Pretty rudimentary.

Done!

Et voila! A finished bike frame.

Ed, me, Yamaguchi, and Haakon.

And the photo my mom requested of June & Koichi.

Yamaguchi Framebuilding School: Day 10

2023-07-10T00:00:00Z

Today really felt like crunch time. We were all working quickly and efficiently. I’m not sure if this was actually to get things done on time or just out of pure excitement.

A few more stories from Mr. Yamaguchi:

Of course he knows City Pop (the music genre)! He used to listen to it while training in the velodrome. It’s more popular now than it was in the 80s in Japan, although its popularity is more US-based these days.

When 3Rensho was a small framebuilding company, they’d occasionally have high schoolers as helpers around the studio in exchange for frames.

Also in the early days, the studio was so small that workers often had to crawl on the floor to get to their stations—while the others were brazing.

After Mr. Yamaguchi designed the Half-Life bike for Valve, they took him on as a sort of representative / translator / intermediary for Japanese correspondence. He was compensated with an ever-updating Steam account with nearly every game released on the platform. He thinks he has around 4,000 games right now. He told me all this after I asked him if he knew Superhot (the game)—he says “super hot” in regards to brazing quite often.

At some point I asked about custom seatposts. They’re not too hard to make; almost the same as a stem. The primary difference is that it’s much more approachable to use two bolts for the clamp, and there needs to be an additional tube to hold the saddle rails (seen on the top post).

Fork crown

Mr. Yamaguchi had me start today by finally brazing my fork crown.

Importantly, we drilled these 3mm holes to line up with air slots on the crown. This differs from crown-to-crown, so it’s important to check.

Flux’d.

After tacking—as usual—we need to penetration braze. For the steering tube-crown junction, we need to heat up the crown body before adding bronze at the top, where the two meet. It’s very similar to how I brazed most of the thicker lugs, but requires a strong flame because of how large it is.

We heat and braze until the bronze goes past the brake holes and starts to be visible under the crown—just like when brazing the seat stay plugs.

Done! Finishing up the dropouts while this cools.

🧑‍🍳😙🤌

Yamaguchi had me file the steering tube until it sat flush with the crown. I didn’t ask, but it kind of seems like this is more of an aesthetics/perfectionism thing than anything structural.

Giving the steering tube the same treatment as the dropouts, making sure to get rid of any and all excess bronze.

Measuring the blade length. Mine should be 360mm.

No need to do any fancy trig here—if we trim off exactly the difference in length, the length will be a little longer at worst.

I took off exactly 14mm and filed both blades to an even height. Sure enough: 361mm. Perfect. (Mr. Yamaguchi said that 1mm extra is fine/good for touring, but 2mm would be too much. If he were building a track bike, he would be a perfectionist and make the measurement spot-on.)

Added air holes to the base of the blades. Inside and at an angle for aesthetics.

Using Mr. Yamaguchi’s custom-made jig to line up the fork, make the blades even, and hold everything in place.

Brazing the crown lugs is exactly the same as the dropout lugs done thus far—tacks at either peak and then penetration brazing between and cleaning up the lug line. However, before penetration brazing, it’s best to close the gap between lug and tube using a pair of pliers.

I finished one blade before lunch…

…and the whole fork just after.

BB shell fillet braze

First, I cleaned up all the carbon build-up from penetration brazing…

…and then just went for it. This small fillet took me an hour and a half, according to Mr. Yamaguchi. He said this was fast.

Haakon admiring my brazing skills :)

I’m proud, okay?? I asked Mr. Yamaguchi if brazing of this quality would get me a framebuilding job. His answer was a little cryptic, of course: you usually start out as a sort of assistant in the workshop, but newbies usually don’t know how to braze. I can show them that I can by building a tower, and then might be able to help if it’s nice enough.

Showroom

The second, and perhaps more impressive, of the Yamaguchis’ bike showrooms is downstairs from the first and in the foyer of their home.

Seatpost clamp

Not much to say here… the process was exactly the same as the stem clamps.

Front disc tab

Mr. Yamaguchi of course has another custom jig for disc brake tabs.

My fork isn’t exactly made for the pressure of a disc brake, so I would have to use a smaller rotor (like 140mm). Additionally, because of my dropouts, it has to be offset in a somewhat awkward position. Once I noticed that I’d have to braze directly onto my beautiful lugs, I started to have second thoughts. We talked about it for a while, and I even had him bring out his testing bike, which has a rear disc tab and no front one (it originally had a disc brake-compatible carbon fork). I consulted Ed too, and both said they wouldn’t do it themselves. I bailed on the front disc. I’ll still do it in the back. The components will be cheaper anyway :P

Rear brake bridge

Setting up the frame in the jig. I noticed Mr. Yamaguchi held it in with only the bottom bracket mount.

The bridge should sit with its bottom at exactly the length of the fork: 360mm. We used the square doodad to mark this location on both seat stays.

Not that there couldn’t be, but there wasn’t a template for this miter cut. It’s a scary one. You basically just line up the file as best you can and go at it, counting strokes, before flipping and repeating.

While I was filing, it got terribly off-center; Mr. Yamaguchi fixed it and it just happened to be at exactly the height we needed.

Using the L-inator to L-ine it up. Of course I sanded, alcohol’d, and flux’d it before brazing.

Four tacks: one on each side, for each seat stay. Then out of the jig and penetration/fillet brazed while on the bike stand.

Fillet braze done!

Super secret tubing 🤫

This bike, which has been sitting above Mr. Yamaguchi’s granite alignment table unnassumingly since we arrived, had a production cost of $6,000,000 (six million US dollars). The US government had developed an extremely light aluminum composite with a stiff outer coating for use on the ISS, and wanted to see if it could win them the Olympics too. They brought a piece of ISS tubing to Yamaguchi’s workshop, but he said it was too large, hoping it would make them leave. They asked for the blueprint and promptly brought him two full bike tubing sets. He built the bike, which is remarkably light, especially considering that carbon fiber wasn’t really in use at the time.

It wasn’t used because Huffy was already the team’s sponsor that year and the material hadn’t been around long enough to test for failure. It was only a few weeks until the races.

Parts of the extra tubing set, never used, in my hand. Mr. Yamaguchi estimates that at the time it cost around $1,000,000 (one million) per foot of tubing.

The seat stay tubing feels—genuinely—barely heavier than a feather.

Yamaguchi Framebuilding School: Day 9

2023-07-10T00:00:00Z

I was pretty tired this morning; I didn’t sleep a ton (stayed up researching bike components!) and didn’t leave the hotel until 8:55. I got to class five minutes late, but woke up in the afternoon with a matcha & my hyperpop playlist :)

Chain stay alignment

Mr. Yamaguchi had me file my chain stay miter cuts some more to get the angle right. My bike is a touring/gravel style frame–I’ll be using medium-large tires–which meant the chainstays should only come a few millimeters from the edge of the bottom bracket shell.

We used allen keys to measure the distance to the BB shell’s edge, locking in the dropouts and using a handmade pressure plate to hold the chain stays in position.

I wanted the freedom to use a 2x or larger-toothed chainring, but the tubing is a little too wide and could cause interference. This is why bikes often have a crimped chainstay section. We measured 40mm from the BB shell and made a mark before pressing the tube in a custom machine (Mr. Yamaguchi built it out of a motorcycle headset installer).

The crimped tube comes out to a diameter of ~13.5mm; 2.5mm smaller than its original 16mm.

We also marked a spot for the air/drainage hole (approx. 30mm down). Mr. Yamaguchi made this in the drill press. It’s important to do it farther down than the dropout is inserted; otherwise the hole will be covered.

Using a square device to align the chainstays vertically. We can also use this to make a sharpie mark at the same distance on both tubes.

Chain stay tacking

We can dip both ends of the chainstays in flux, since it’s liquid-y today and won’t gunk up the opening. The holes are slightly too small to put the brush in.

Using the L/square device to align the dropouts. This is a pretty exact science that Mr. Yamaguchi took care of for me. If the dropouts are off in any way, the seat stays will end up lopsided. Down the line this can lead to a squeaky brake since it will touch the wheel unevenly.

This model of Campagnolo dropout from the 80s was wildly popular, but made 3Rensho double check all of their tools for misalignment–they were notoriously uneven. Yamaguchi found that even larger framebuilders like Serotta and DeRosa were cutting uneven seat stays to make up for the poorly-cast dropouts.

Tacked chain stays, BB shell end.

We can leave the allen keys in for alignment, as long as I don’t accidentally braze them on. We left the two outer tacks for when we got the frame out of the jig (post-seatstay tacking).

Tacked chain stays, dropout end.

This was a somewhat tricky tack. We need to warm up the dropout more since it dissipates heat through the jig so well. Mr. Yamaguchi taught me to use the the flame extremely close to the blue cone to warm up just the corner of the dropout to keep the tubing from burning.

Seat stay prep

Mr. Yamaguchi cut the notches for me this time… Maybe he noticed I wasn’t too good at it.

Seat stay length doesn’t need to be exact, as long they make solid contact with your seat tube (or top tube, for triple-triangle). We just held the plugs where we wanted them to lie, marking the tubing around where the plug would insert.

Making sure the two seat stay cuts are level.

We need to braze the plugs from the inside out because of how narrow the gap is. To do so, we create a small bronze ring, just wide enough for the tubing opening, and push it into place with the plug.

The ring, in position.

We hold the seat stay upside-down with this magnetic welder’s holder device.

Mr. Yamaguchi explained to align the plug using a straight edge at both ends. He does it visually, but I figured it would be most exact if we used a yard stick for alignment.

Brazing is done by heating up the plug on all sides until it is cherry red; bringing heat up to the ring; then bringing the bronze back down until it just starts to flow out of the tubing.

Me brazing the second seat stay.

Sanded seat stays being positioned on the frame jig.

Again aligning the seat stays before adding air holes.

Marking the air hole positioning. Same as the chain stays.

Tacking the stays to the seat tube and the dropouts. Also same as chain stays.

Removing the dropout jig.

Adding a 135mm axle for maintaining spacing while I braze.

Finishing the chain stay-BB shell tacks.

My final frame geometry! Obviously there’s more brazing to be done…

Penetration brazing the chain stays to the BB shell.

Brazed.

The video stopped short, but this is me doing the final brazing on the dropout end. The notch edges need to be filled in for each tube (4 per tube, for a a total of 16 edges).

Filling in the end of the tubing is a little difficult, but generally we just heat up the dropout until the edge of the brazed section starts to melt again; then the bronze brazing rod is just fed into the gap from one side until it closes up.

Done!

And finally, fillet brazing the seat stays to the seat tube. Not so difficult, but symmetry is a pain in the ass and I’m a perfectionist. This took a while.

Somehow the three of us had moved around in such a way that we were all brazing in one corner of the workshop. While maneuvering between us, Mr. Yamaguchi stopped to check my brazing and I accidentally burnt a hole in his Team USA t-shirt. He was way too friendly about it; he must not have liked the shirt that much 🤭