sottol

> Sometimes I really wanna skip 10 years into the future and see where all this eventually leads.

Sounds glib, but honestly ... more Ads :/

But WHY do these agents even want to communicate (specifically in the first openai incident, there seems to be instructions to communicate here)m

It must be some artifact of their RL or training? Maybe it's RL for agentic workflows with planner, coordinator and workers?

Does anyone have some insight on this? I'd be very curious.

You did say "consumer tech" - so you're probablt aware of the BT-46. But this has been done a few times before... though it's probably more dangerous than ever with up to 1MW (1000kW) output from electric motors. Some of these were actually sold afair.

I'm not sure about those thrusters though, I doubt they'd be legal?

[0] https://en.wikipedia.org/wiki/Brabham_BT46

[1] https://en.wikipedia.org/wiki/Gordon_Murray_Automotive_T.50

[2] https://en.wikipedia.org/wiki/McLaren_F1

[3] https://en.wikipedia.org/wiki/McMurtry_Sp%C3%A9irling

It feels like pulling and building someone else's code. It just happens to be *exactly* what you're looking for.

This is the best way I can describe how I feel about it.

When the software output is the _goal_, it's a lot less rewarding. When it's not, it's fine I guess. I'd still feel better if I had it done it myself but it's fine.

Semi-related, but I did recently wonder if China's real-estate bust might serve as a strong fertility support measure / incidental pronatalist policy.

Tier-3 and tier-4 cities supposedly saw 30-50% property value declines, setting them back more than two decades. If real estate was really holding back families, as is often stipulated, this might reverse that.

Otoh, rural Japan has lots of cheap or even free real estate but it doesn't seem to support fertility rates at all.

I just got my PCBs - yay! I will try soldering the switches and MCU this weekend or early next week.

I'll try to write up a post if I get it all working. Probably even if not... publish your failures and all.

Thanks :(

I'm still waiting on my order of 15 PCBs done with Sol - a mini-BT-keyboard that I hope will be comparable to the Psion 5 / Gemini PDA ones, using Kailh PG1316S switches. I still need to design and manufacture a housing (I do have a lot of hobbyist experience with this though), but the idea is to have a magsafe-connectable keyboard I can clip onto my Pixel and make use of the new Linux subsystem (or Termux) and do some productive stuff on the go. It's just been manufactured by JLC but not shipped yet. Funny enough, production itself was ~$70, then $80 for shipping, taxes and fees (tarriffs?).

Instead of directly generating Kicad, I chose to let it write "generators" in Python that output Kicad (S-Exps!) pcb and project - with mixed results - but I can generate a version with/without per-key Neopixels and customize certain parts of the board, including debug-stuff. I also _think_ this makes it easier to resume working on this in new sessions but not sure. Otoh, I never actually looked at the Python (vibed...) so it's still a lot of hard-coding.

I also landed on and used KicadRoutingTools after trying FreeRouting, combined with some minimal "auto-placing" code for larger ICs and such that are placed with certain constraints. My routing runs are way slower though - on the order of minutes.

I always wanted to try atopile - that might be a good match for AI/PCB work. The only thing that really held me back was the lack of routing.

Anything that puts pressures on Nvidia's margins is going to have an outsized effect imo - margins declining from 75% to ~50% would undo all of its projected 2027 growth of 70%.

And what about the purported "AI frontier slowdown" - if frontier labs stop pushing better and better models, the main way to grow the pie is more users and that will eventually stop, too. Then providers would probably stop buying new hardware hand-over-fist and move to a slower depreciation-based replacement.

Thanks! Qwen-3.8 27B seems to benchmark better but I'd like to try this some time.

The press release: https://ifm.ai/k2/press-release/

It's not the blog post, but there's some info here:

https://ifm.ai/k2/

375 A23B, 36 A4B, 32B, 7B, 3.7B, 0.9B variants.

> 32B: Ranking among the top models in its class, 32B is our most powerful dense model, balancing capability, adaptability, and local deployability.

> 7B: The industry’s best-performing model under 10B combines strong software engineering and expert knowledge in a package small enough to run on a phone.

I'm torn on this - on the one hand performance matters, on the other so does capability.

I could run Qwen 3.6 27B on my laptop, but at 5 tok/s it was too slow even without overthinking - I never used it. OTOH, Qwen 3.6 35B A3B ran at 20 tok/s but it just could not get done what I asked of it. It sort of got close but you had to repeat and retry so much that it might have been faster to run 27B dense... maybe?

So that said, I might take a much better model that runs 2-3x slower (total time per task) but that's more capable over a faster, less capable one.

I'd also like to try a proper "plan-then-execute" type execution where thinking is entirely disabled (or low) during the execution stage but enabled/max during the planning stage.

I will definitely give 3.8 27B a better shot than 3.6 though.

A lot of the benchmarks seem often near meaningless these days - really bench-maxxed to the hilt. I tend to still look at the Artificial Analysis rankings to get at least an idea on relative performance of models, is that still warranted?

What or other opinions on how representative the AA rankings are of real-world performance? Any better indicators?

There's 5B working age people world wide and only 50% of those make more than $500/mo where it's even conceivable to spend $50/mo. I did the napkin math, it can still sorta kinda work out.

I think more realistically we'll have something like the Google/social media US-vs-world profit split of 40-50% US vs rest of world combined. Even those numbers can work out but then I don't see tremendous growth.

But the sector valuation is already priced for wholesale workforce replacement or massively expanded productivity and AI platform providers taking a lot of that pie for themselves.

With corporate profits already near all-time highs with respect to GDP, who is going to buy all those new products (from expanded productivity ) if all the gains only go to OIA and Anthropic employees?

It's an interesting time.

Most of them live and breathe biking. They choose to make close to minimum wage to "do what they love", they don't own the store. I think it's genuine excitement and enthusiasm that you take as elitism. Often, it's a big part of their identity... "how could anyone else _not_ care as much"?

I used to bike with a few bike shop employees and they were only in it to eke out a living doing something adjacent to the thing they love, selling bikes was not their passion, riding them and talking about the tech was.

At least 10 years ago, wages in the whole sector were pretty low, so you'd only work there if you really wanted to and valued perks like buying bikes at wholesale (hey, that's half off) and having access to tools or a mechanic buddy.

BMW did have the _iconic_ "halo angel eye" headlights in the mid-2000s - the most iconic headlight signature at least in my memory. So your point would make sense.

Those were often copied by enthusiasts and "ported" to other cars via aftermarket headlights.

Afair, e.g. Land Rover stopped producing the classic Defender due to more stringent pedestrian protection regulations in the EU. They introduced strict front-end safety requirements for cars and SUVs whereas the US does not. So vehicles designed and sold in European (in numbers) are probably safer for pedestrians because it is tested for and required - but there are exceptions for some US trucks somehow.

> You just made that up. Pedestrians are mostly killed by head onto solid object.

Not to be combative, but I'd also like to see stats on that - that sounds just as made up. I'd expect a lot of pedestrians to strike the hood (about just as likely as windshield) as most pedestrian accidents happen in parking lots, drive-ways, traffic lights and vehicles exiting across a sidewalk (under 25mph).

Hard to find reliable data, but could only find (shared) data for 2022 for large SUVs - whatabout doesn't quite cut it imo, feel free to look into midsize SUVs but imo the really large ones are more problematic:

Ford Expedetion: 81,988 in 2021, 71,648 in 2022

Mercedes GLS: 24,482 in 2021, 12,395 in 2022

Chevy Tahoe: 106,030 2021, 109,032 2022

Chevy Suburban: 48,214 2021, 52,459 2022

Dodge Durango: 65,935 2021, 58,627 2022

GMC Yukon: 84,242 2021, 80,731 2022

Nissan Armada: 22,814 2021, 17,551 2022

Toyota Sequoia: 8,070 2021, 7,066 2022 (but about 25,000 after refresh in 2023)

* Ford F-150 [2] *: 726,004 2021, 683,633 in 2022.

The Large SUVs are all linked here: https://carfigures.com/us-market-segment/large-suvs

I'm sure the mix is different and skews a lot more Japanese for mid-size (which are probably also dangerous), but these large SUVs and trucks with huge hoods and no rear visibility are getting quite problematic [1]. The front blind zones are pretty ridiculous, to the point that people advocate for front cameras: https://youtu.be/NDH3FDfVQl0?si=o0uvSEmKZrn2gHSN&t=205 (with timestamp).

[1] https://www.nbcnews.com/news/us-news/americas-cars-trucks-ar... [2] https://carfigures.com/us-market-brand/ford/f-series

There's Will Whang's boards - IMX585 [1] (16:9 1"-ish) and even IMX283 [2] (1") and even IMX294 [3] (Micro 4/3). But just those camera boards run $199 to $399, and released in "artisanal quantities" (I think their hand-assembled!)... so you have to pounce when restocked. Soho Enterprise has some IMX585 boards as well and I've seen some IMX585 MIPI CSI boards on aliexpress afair but never tried them

I'm experimenting with and have built a rangefinder-style camera [4], built around the IMX585 or IMX283 (the only boards I got my hands on) but using a CM5, this thing gets hot. It works though! Not too much bigger than my Leica Q. Haven't released anything yet but I tend to work on it and the model is in OnShape. Currently planning a complete screen-less redesign in FreeCAD... so that's _really_ different and slow, but I'm so over proprietary software :/

There's also the CinePi project using those sensors on a full-size Pi with a pretty active discord server.

[1] https://github.com/will127534/StarlightEye

[2] https://github.com/will127534/OneInchEye

[3] https://github.com/will127534/FourThirdsEye

[4] https://cad.onshape.com/documents/29c9488b2d4b80b73bcf3980/w...

I don't think this is solely tied to car manufacturing automation. Even though Hyundai Motor Group is acquiring them, I would imagine they'd be well-positioned to commercialize general-purpose robotics and not just for car manufacturing, if Tesla is anything to go by.

I do think this might be tied to South Korea's demographics, by 2040 the working-age population is projected to decline 25% from 2020 and keep declining almost linearly until leveling out around 17M around 2065, a 50% drop total in < 50 years. I would think HMG / Hyundai sees a huge business opportunity or this might be a national-level political priority but I don't know the specifics.

I've sort of been thinking about this as well. Personally, I'd like to re-capture the era of personal blogs and niche knowledge discovery of the earlier web I experienced - crossed with something easy to host/publish and not requiring a browser.

I don't really have any coherent picture but I would like to see these ideas I think:

- Anti-commercial/anti-tracking: maybe requiring some sort of open-source license for all published content that makes it harder to commercially exploit the information, ideally this would be by and for the community, especially in light of recent aggressive LLM-training crawling. I would also like to exclude advertisement and tracking.

- Browser-less: The idea would be to do away with the complexity of the modern web (as people say, browsers are basically operating systems), back to more of its hyper-text roots. Simple documents, mostly textual information. I could imagine a mix of basic markdown and some pre-wired complex/interactive views like "forum" or "blog" and so on (differences in how data is loaded, presented, ...) - the idea would be to implement the "app" part in the browser-replacement and not in the web-page itself if that makes sense. This would lead to more uniformity but that might be a good thing. I'm not even sure if/how images would fit in or videos.

- Peer-to-peer?: Hosting should be as simple as hitting a "publish" button on an article. I like the idea of decentralization, so maybe there could be some sort of peer-to-peer federation where users could "host" content that they've read, liked or general content that's part of a certain (sub-) community. This might require some ranking like HN or a similar mechanism to (unfortunately) censor certain content if the community would not believe it to match their values - so not ultimate freedom. P2P would be more about decentralization, and maybe anti-tracking than pure censorship-resistance.

A session might look like opening the "non-browser" app - it would be fast and require very little memory. Then you'd select or type a community/site and you view of all the content with filters and sorts, depending on the community/site's "template" (again, this is not JS/HTML - basically a native form rendered directly if you will). When you feel like it, you click the "create" button, a text-area + preview pops up and you write your post or article in markdown. When happy, you "publish" and it gets slowly disseminated through all the P2P nodes of your community. This could encompass communities like HN or reddit even if the voting mechanics are worked out, personal blogs, ... but would probably exclude e-commerce stores or video sites because the engine would be potentially too simplistic - and that's fine by me.

But these AIs need energy and GPUs data-centers ... who pays for those? I could imagine a circular mini-economy between a few companies making the bare essentials to keep AI running and not catering to 99.9% of the population because they don't have the funds to buy anything those companies could produce so they don't.

In that scenario AI and robots produce everything, the owners of those AI companies can trade their AI's output with other AI/robot companies, robot and chip manufacturers and commodity owners? So 10000 people world-wide are fine, everyone else is not?

Exactly. But I did find an article ([1]) and spend doesn't seem that high per engineer ($150 to $250 per eng) - at least on average, I assume the costs were skyrocketing towards the end.

> Adoption climbed from 32 percent of engineers in February to 84 percent classified as agentic coding users by March. By spring, 95 percent of Uber engineers used artificial intelligence tools monthly, and roughly 70 percent of committed code originated from those tools. About 11 percent of live backend updates were written by agents with no human in the loop, according to Uber's own disclosures.

> The numbers behind the spend are what make the story instructive rather than anecdotal. Monthly cost per engineer ranged from $150 to $250 on average, with power users running between $500 and $2,000.

My guess is that the reason to rethink AI-spend was probably the exponential growth in cost over time, and tokenmaxxing payoff not being immediately obvious as mentioned in the article.

[1] https://www.forbes.com/sites/janakirammsv/2026/05/17/uber-bu...

Maybe that's the plan :)

But on a more serious note, do we know how much Uber spent per technical employee/month? I assume it is far more than even any of those $200 "max ai" plans.

And the other question is how much the public would be willing to spend, in my estimation this is as "cheap" as it will ever get (main-stream at least).

Otoh, you can buy contact lenses OTC in other European countries... not saying Europe isn't over-regulated (try legally building a bike-trail on your own forest-land in Germany!! ...insanity) but it's a bad example.

Once you have a reliable printer, the workflow is mostly to slice -> send to printer -> wait and check on it every couple of hours until it's done ime. Imo it no longer super matters how much better the on-screen ui or webcam are.

Mutli-color though is where Bambu has a good leg up.

(Diluted) Vision Miner Nano Polymer Adhesive and a good bed leveling probe has done a lot to make my printer set and forget, no matter which print sheets I use.

Bambu also tried to patent several widely used techniques in china, fyi.

https://www.mdpi.com/2411-5134/8/6/141

My thing with bambu was always that they polished whatever the industry (and hobbyists) had invented and closed it all off, then also innovating on top of that but never giving back unless they _had_ to. Polish and mechanical design are great but corexy kinematics, input-shaping are imo what made the X1 stand out as the fast+good-qual printer when it launched. A lot of what they added on top was then to build a moat.

This may be a controversial take, but imo it would be Bambu to set the industry back by a decade if they "win" and lock up the market. That's clearly their strategy afaict.

Does anyone remember Bambu patenting existing open inventions as their own? I can't seem to find good links anymore (?!) but there's some details here https://www.mdpi.com/2411-5134/8/6/141

I would have said Prusa a year or two ago but they've reneged a little on their open-ness. That was probably in response to Bambu being fully closed and gaining so much market share.

The Core line of printers seems promising and a big leap towards closing the gap towards Bambu's corexy printers but haven't used one yet and I've been out of the game a little. Bambu though is probably more of a high-end appliance type than Prusas more utilitarian feel.