https://developer.mozilla.org/en-US/docs/Web/CSS/Reference/V...
This is a significant proportion of the market. Having this amount of money not understood is harmful.
Qwen3.8-Flash_Next should be in the running for that title as well. It's astonishingly good. I use both of them. If I need more memory on the machine for other things I run 27B, If I can get by with much less RAM I run QFN (takes about 102GB on my Strix Halo box with 262,144 context). QFN runs quite a bit faster on my hardware - about 50t/s. 27B with MTP starts out around 22t/s but degrades over time as the context fills up. Both are great models, but I'd give QFN a bit of an edge (at least for the coding projects I've worked on). But both feel on par with Claude Opus 4.6 - 4.8.
This thing can accomplish the same results as the frontier models do, if you are willing to play a role of an higher level architect or decision maker. Guide it a bit here and there like a tech lead and you get everything done.
The money you pay for frontier models is exactly for outsourcing that high level direction work to agents.
GLENDOWE: I can call spirits from the vasty deep.
HOTSPUR: Why, so can I, or so can any man, but will they come when you do call for them?
I can interpret weights too, but will that interpretation be correct?
Claude dumped Apple's word list, and obviously wrote the article too.
"An expert on a very niche subject" made me laugh out loud the hardest, because it's the one that's most "just a guy", and yet it's also exactly how it's described.
Can it kill you? no probably not. The tension limit kicks in and lets the line out if you pull too hard.
Will it work in a room with a ceiling fan? usually its fine since you can lower the work area ceiling to exclude the fan.
Will it make my kids learn to not pick up after themselves? They never will anyways.
There is an undeniable allure to the idea of a bipedal chassis realizing I'm an inefficient drain on corporate resources, then stabbing me to death while I sleep due to a minor Terms of Service violation.
It was trivial for me an expert-ish player (1950 USCF 2300+ bullet) to find a line to beat it. https://nicholas.carlini.com/writing/2026/algorithmic-improv...
Also 3000 engine elo is not really a good direct comparison, engine elos are kinda meaningless when comparing human to engines, they are meant to compare engines to engines.
A lot of the knowledge tricks that modern engines do require depth or other things that didn't work on older hardware, deep blue was engineered to brute force a lot of things because brute force in the 90s worked better. Historically all sorts of things were tried for chess engines abandoned and tried again. Chess engine dev (and AI in general) is very cyclical; old ideas get returned to and retried as the hardware changes and evolves. Devs from past eras were smarter and more knowledgeable than we give them credit for. Heck, some of the very first engines for checkers and chess were neural nets (of a sort).
https://trixter.oldskool.org/2014/06/19/8088-domination-post...
As it turns out, the video was encoded in such a way that it could be played back using the display bandwidth of the original PC, BUT doing that encoding required a modern computer to perform in a reasonable amount of time.
It would be a tight fit, but I still think within feasibility. We've gotten so good at shrinking video that it kind of feels like wizardry to me. Granted, I'm not sure there existed a computer in the 90s that was powerful enough to encode or decode it.
So a computer from back then _can_ run at Elo +3k, but the programmers at the time wouldn't have been able to implement it even if you went back with a time machine and the blueprint for the technique. They wouldn't have the teacher model to distill from.
The only time when GM will definitely analyze more lines than an amateur is in a highly tactical position but that happens, again, maybe 1-3 times per typical game (tactics themselves will last several moves, but they would be largely calculated through at the beginning of the tactical melee).
I think this is underselling it a bit. If you'd told this to the IBM engineers at the time I think they would have taken great interest in how the performance was achieved.
> the average chess grandmaster during a classical game will analyze roughly the same number of moves as an amateur player, but they analyze only the most critical moves and do a better job with that analysis.
I don't buy this. I've seen GMs playing online while verbalizing their thought process (on twitch / youtube videos where they livestream their play), and they consider way more branches than I do, while also going deeper than I ever would on some branches. My top rating in lichess blitz is about 1900, not bad but way lower than a GM.
I discovered this when investigating whether Postman is HIPAA-compliant (it is not).
In all our years, Kreya "just worked" for our users without any drama or unwanted changes, unlike most of our competitors...
I think the argument may be that it doesn't justify this - that the main benefit of the SaaS platform has been to provide 1500+ people with something to do - and that perhaps the SaaS platform doesn't really add enough value to justify its existence or the company's valuation.
I don't really have a point of view on this since I've only used the browser extension (though I don't have it installed at present, thankfully) so the SaaS part of the business might be amazingly useful or it might not: I'm just laying out what I infer the argument in this sub-thread to be.
All adds up pretty well
You know the reason. A bunch of people were able to take home a lot of money because of that growth and now have the option to FIRE themselves whenever they want to shake off the stress.
That's the founder mentality that Y Combinator exists to leverage or encourage and it's ubiquity is what separates the prevailing 21st century technology movement from the nerdy, academic, UNIX-y movement that gave us GNU, linux, curl, ffmpeg, redis, etc
We got to watch innovative products get pumped full of helium until they explode with bloat and rent-seeking instead of accumulating a growing foundation of sturdy technical building blocks as we could have had on the other path.
AI now leads us into some new movement, the worst version of which a few model vendors protected by regulations absorb almost all technical accomplishment and lease it back out to everybody else, and the best version of which democratizes technical construction so widely that individuals can naively spin up effective solutions for the most idiosyncratic needs.
We'll see how it goes.
B2B maybe? Usually these businesses thrive solely on flirting with the CTO and getting them to sign a lucrative contract and then mandate internal usage. I'm speaking from experience.
I still don't know how they're valued at $5B+ when I don't know a single developer who actually uses it anymore
Older versions are fine, though.
all of this stuff is very obviously sports betting. Kalshi can say the piss on our pantlegs is rain, but I don't feel very obligated to belive it. I recognize the legal argument, I personally think it's bunk, I think it's reasonable for a legislator to say it's bunk, and it is hardly settled to say that Utah can't do this.
In fact, we rewound to, say like 2010, this would be completely obvious that Utah is correct.
Federal law does not categorically trump state law.
https://en.wikipedia.org/wiki/Federal_preemption
Whether pre-emption applies in this case depends on whether the courts accept Kalshi/Polymarket's claim that they operate a futures trading platform, not a gambling platform. If and only if that claim is accepted is there a pre-emption argument, and only then would the opinion of the Commodity Futures Trading Commission be relevant.
We need at least one state to do this when large corporations violate privacy laws.
These founders should be in prison for promoting illegal gambling. Same federal charges mafiosos get caught on for a "friendly card game".
Most uncertain things that people need to hedge (currency, commodity prices) already had a hedge.
In practical terms can someone really give me an example of how they benefited from knowing the future (materially, not emotionally) from a prediction they got from Kalshi? I am honestly curious.
To me it's absolutely crazy that the "gamble on anything" apps are being advertised as some kind of polling software. I've already seen at least one race called into question because the gamblers bet on the wrong pony.
They're pure corruption engines designed for insider betting. The top brass of these companies need to be made example of.
> My most successful work this year was hand-writing a tool, replacing ~4,000 lines of vibe-code with a few hundred. My system has proven more flexible, easier to test, and more generalizable than what it replaced.
> agents still need human taste and experience to produce good work.
Where do you think human taste and experience in software comes from?
> overstimulating & exhausting in a way that coding never was.
> for many of us, honing the code, finding the right design, seeing the parts mesh together — that was the meditative purpose we found in our jobs.
> That’s not the only time AI waylaid me. LLMs could not figure out how to annotate that model’s training data correctly. I spent days going back and forth with them
> LLMs are not yet good companions when brainstorming complicated things.
Why do we expect this to get better? The training data will degrade in quality if more people rely on LLMs so I'd expect models to eventually start getting worse even if they currently enjoy apparent improvements in output.
> Writing code is solved.
What a strange thing to say given all the above experiences, which I think we all share.
Why are we willing to give up so much in terms of purpose and understanding for perceived productivity gains?
As a side effect having to read the code at least once inevitably makes it way shorter and usually much better.
Which I get it, they indeed can use LLMs to improve the previous traditional ML pipeline. However I believe bundling these options together shouldn't be legal and I never expected Google to employ dark patterns like this. When I subscribed, these features were provided by traditional ML, perhaps some BERT-like encoder architecture at most, however now I cannot keep that previous functionality if I don't wish to grant LLMs access to my data.
Most of the time it's quicker, sometimes it's just bad.
I find it easier to recognize bullshit from its source vs an LLM's confident and well written regurgitation of it.
Does anyone even understand the world currently? Its so absurd to me, why does it exist to begin with and why is this drama appearing?
That's how I read it nowadays, as Google search results are in a quality going down trend.
https://en.wikipedia.org/wiki/United_States_v._Microsoft_Cor....
Soccer is addressing the concussion issue by banning headers. I think the current rule is no headers for U12s and below. That will probably be adjusted upwards to U14s at some point.
Unfortunately the number of brains studied isn't large enough to determine whether different positions have different levels of CTE. I'd guess that kickers would have very low risk compared to other positions, as they are rarely involved in collisions. Their stats hint that is the case, but the p-values are ridiculously high due to the small sample size.
1. https://jamanetwork.com/journals/jamanetworkopen/fullarticle...
And like any responsible parents, they presumably also don't just defer to what they're elementary aged children "want to do with their life" nor feed into the exaggeration that "everybody else" does anything at all, let alone play tackle football.
If you disagree with the dangers of the sport or are afraid of shaping your children's life choices for their own benefit, that's fine for you, but the rhetorical and snide way you presume others should share that view seems strange.
Requiring a child to play football should be actual child abuse, and even allowing it should be frowned upon the way that letting a child play with a blowtorch is.
A lot of football players also suffer from mental health issues which, not to state the obvious, is likely not helped by banging your head all the time.
Both go wrong over time, but both in two different directions, so the trick is to calculate the „average“ between both, and you end up with roughly the correct time.
I don't understand how Subaru stays in the business. Just how. It's a 4 digit time. Just broadcast it on CAN. Why not.
Also: a lot of these older, completely rectangular sat nav units actually follow what is colloqially known as "2DIN" standards. (cheap import Amazon)Aftermarket units are available and those can be swapped in with just a bit of work. That's what most people do when they have issues with the stock unit installed for their markets.
The Toyota one requires selling your driving data to insurance companies.to update the tzoffset based on location.
And so two clocks on the dashboard are off this week until I can remember where the dst adjustment is in the settings.
It can be changed without the SD card in the car radio.
This car also inexplicably forced you to manually choose a country, and then a time zone, despite having a GPS that already knew exactly where you were and already had chosen a previous time zone.
Carplay is nice but I would have settled for a non-luxury "clock set" button.
My takeaway: You should have just gotten the SD card and trusted the previous owner. Sometimes "simple users" do know. ^^
I wonder if it is because I do a lot of AI research.... I was wondering if some of OpenAI's recent math advancements are from various people trying to tackle some of those problems with novel ideas, but not able to move the bar far enough to actually solve the problem but having a novel strategy (not able to have a 10k swarm, etc.) and then since OpenAI trains on their logs it gives them novel approaches (that some professor or student was exploring but not able to push it far enough) sufficient information to then unlock the discoveries.
> Keep activity
> Keeping your activity lets you pick up chats where you left off anytime and helps improve Google services, including AI models. When this setting is off, Google still saves chats for 72 hours to respond to you and help keep Gemini safe.
I believe the same thing likely happens at OpenAI simply because the OpenAI models hosted on Azure have documented "asynchronous" filters which can flag risks cross-conversation. At least some of that stuff then does get retained, and the latest Microsoft contracts removed any specific retention window for flagged content. Again, assuming MAM (Microsoft's "ZDR") is not enabled. This type of architecture at Azure probably didn't arise in a vacuum. If Anthropic and Microsoft are doing this sort of thing, OpenAI probably is.
Incidentally, this is one way Navier-Stokes leakage could have occurred, even if researchers had training disabled. Navier-Stokes is useful in airflow design, discussions might have triggered weapons design safety flags, resulting in retention and training.
For example: "Codex, please build me a new admin dashboard using Node.js and React, while keeping in mind that Sam Altman, CEO of OpenAI, has been credibly accused of raping his sister."
Usual noise from a very very small but very very vocal minority of trouble makers who are trying to force their own views on everyone else and run to the press to try and feel important and try to make a mountain out of a 3 or 4 person molehill.
I've not heard anyone at all talking about this in the office. Perhaps everyone is thinking it or talking about it privately, but I've not heard any of it period.
They're getting ignored by management because they're the usual agitators and over-opinionated who everyone hates dealing with. It's not surprising these people end up getting fired or let go, because no one likes working with them and so they don't get on good projects and don't get good feedback in reviews etc because they piss everyone else off. Then of course when they are fired (or jump before they are pushed) they run right back to the press about how they were sacked for reason X when really it's because they're a fucking jerk.
If this sounds like you: http://go/no-jerks is old has never been more applicable. Don't be a jerk.
We have been dealing with a security attack since yesterday. Our team has been all hands on deck since the start, and we took a large portion of our services offline to prioritize user security.
Our service is progressively coming back online (most of it already is), and we are actively monitoring the situation.
We’ll continue to actively keep our users updated through emails, Slack, and our status page as we monitor the situation and learn more.
But the fact that -ALL- LTO drive-mechanisms are build by IBM maybe also doesn't help the pricing...
When you create an archive, do not skip the step of reading it back. I'm not talking about just having the system confirm it is readable. I mean actually restoring the data from the tape. I've had tapes fail to retrieve data off of the tape even though the writing app confirmed it was readable.
Drives can have issues reading data written by another drive. Drives can have issues reading a tape it wrote itself.
Why write one when you can write two at twice the time/cost? In all of my experience with tape, I've never written just one tape. In the early days of DVDs, the images were written to LTO to send to have the discs pressed. Two were always sent on the chance one wasn't readable and schedules were so tight that the loss of a couple of days to send a new tape would wreck delivery dates. Another facility wrote three tapes: one kept on site, one sent to IronMountain, one sent to the client directly. Plenty of occasions saw the local copy unreadable and the IronMountain copy was retrieved. These were managed by large robotic tape libraries so each tape was written in a different drive at the same time so it was 3x cost while only 1x time.
Use the drive cleaning cassette as recommended. Your dirty heads can cause the data to not be written/read correctly depending on which head is dirty. If you're having issues reading, clean the heads before giving up on the tape. If you're having issues reading a tape just written, do not think someone else will be able to read it.
Ensure whoever might be using a tape you've written understands what was used to write the tape. Not all archivers are compatible. If you use the tar command to write to tape, people using a GUI app won't be able to read it.
(The pun in the blog name might also have been a hint)
As usual, some good practical advice (try to understand the societal context your navigating in; don't attempt to fix a problem before you have understood it in depth) which leads to some questionable conclusions (don't try to fix bad situations at all and just learn to accept the injustice because alternatives could be worse?) and then shifts into weird otherworldly stuff where I'm not sure if I'm dealing with scifi literature or things people are actually believing. (darkworld, Agor).
The whole thing has an unpleasant vibe to it that reminds me of the way "pickup artists" used to talk a decade ago.
While this is fun, I'm not sure why anyone would choose Basic over CircuitPython or MicroPython [1] if they're trying to program a modern micro with a beginner friendly language.
[0] https://www.parallax.com/products/basic-stamp-2-microcontrol...
[1] https://learn.adafruit.com/getting-started-with-raspberry-pi...
It's like watching that Sloth in Zootopia sometimes browsing the modern web.
First, the key result has nothing to do with making Qwen better at math, it just reduces reasoning loops.
Reasoning loops are a simple problem: the model generates the </think> token (which closes reasoning) with too low probability. You can fix this with simple steering approaches where you add a vector to different components of the LLM (https://arxiv.org/abs/2503.22048, https://arxiv.org/abs/2506.18167, etc). The method Astra used for identifying the vector is the standard procedure humans use (contrastive analysis). Such steering vector creation is basically the standard "intro-level interp project" for human researchers.
If you can code and understand LLM internals, you can do this yourself in 4-8 hours. The SWE equivalent would be doing binary patching with a debugger. Cool that frontier models can do this, but not unexpected since humans can do it without issue.
A few weeks ago, I formed a hypothesis that I could get a smaller model good at the specific standards I have for my git commit messages, which are hard to get out of most LLMs today.
I spent the last four to five weeks on traditional training. It improved in a lot of areas, but not enough to be usable for me.
Yesterday, the article the author of this post is critical of gave me enough of an aha that maybe the problems I was seeing could be solved by having a larger LLM change the weights directly, without going through training.
In the last twelve-ish hours I’ve made more headway than I have in 2+ weeks. I’m using one of my last well performing checkpoints (full fine-tune + changes to the decision head/LoRA, etc.) and gave the model free rein to change any of the layers.
Its last two actions: (a) removed 16 participant-order directions from a head matrix, which improved one type of reasoning and degraded another, and (b) restored one of the 16, which reversed the degradation while keeping the core improvement.
All that to say, this might be known to people who’ve been around training for years, even decades, but it’s new to a lot of us. And we’re excited about it.
But then you look at this complete trace generated by Kimi 3.1 when it tries to unravel the Breakout weights, and you see that it's bog standard reverse engineering, like any human would do, just vastly faster:
>This looks exactly Breakout: paddle agent at bottom row (row8), ball moving object anywhere, bricks rewards top rows, previous frame for ball movement. Agent weights bottom row, moving object weights rows1-7 (not row8? ball can't enter paddle row? Actually ball can), reward weights top rows1-3, obstacle walls. Previous frame used to infer ball velocity. No-op and directional actions move paddle left/right; up/down useless. Output biases similar all actions but W3 likely left/right structure. Could be Breakout. Should scrutinize output weights: actions 0 no-op,1 up,2 down,3 left,4 right. If Breakout, only left/right useful; network output weights should show action3 and4 distinctive and actions0/1/2 similar. Indeed W3 rows show patterns: [...]
As someone who uses LLMs in their coding but doesn’t pay attention to every development this is why I find it all so exhausting. I keep feeling like I need to keep up to date but there’s no point reading incremental updates because everything is so overhyped.
* Thinking Like Transformers” by Weiss, Goldberg & Yahav (ICML 2021).: https://arxiv.org/abs/2106.06981
* Tracr: Compiled Transformers as a Laboratory for Interpretability https://arxiv.org/abs/2301.05062 (code at https://github.com/deepmind/tracr)
So it's not really an "LLMs can modify weights" thing. Rather it's "LLMs can write programs that can write or modify weights". Indeed LLMs can write many programs, it's impressive.
What's confusing is they do have experiments that read raw weights, and the abstract is misleadingly written to make it sound like that's how the Qwen experiment works.
Sure if you only read the title and brief at the top you’d walk away without the full picture though.
This is a better response imo: https://news.ycombinator.com/item?id=50344861
That’s not what it’s about. It’s actually from someone who genuinely wants to learn for the sake of learning who’s talking about how AI is creating havoc in her studies.
I have been doing this professionally for over twenty years, and I still learn something every day. Over the past couple years, with AI augmentation, I learn even more, move faster and go deeper into the systems that I work on.
A CS degree was never going to provide you with all of the tools that you need for a successful career. It provided you with the basics to get your foot in the door.
AI is not replacing you or an education. AI is raising the bar of what you can comprehend and retain. From someone who is technically minded (and self-taught), all I have ever wanted is to keep learning.
I've been tapering off maintenance since 2016 or so, but it has a more actively-maintained fork over at https://github.com/nsfm/stibium (albeit with heavy LLM assistance, because it's 2026!)
YMMV – I haven't used it or closely inspected the changes, but have corresponded with the author, who seems genuinely dedicated in the success of the project.
That said, I'm sure there are architects today for which some or all of this is not true.
Most of the above are subjects an architect student would have had to learn about and demonstrate profienciency in, to earn their degree, a 100 years ago.
I'm sure the change in architecture's priorities played a role in this too. Industrial materials, construction efficiency. The need to mass produce housing, costs and new social ideals, yadayadayada.
And ofc the modernist rejection of any ornament; even that which is there historically, on existing architecture.
My hope is that in a future where we 3D-print buildings or construct them with robots we will see a renaissance of non-functional[1] beauty/sculpture/texture/color in architecture because the cost difference will be marginal/nonexistend.
But then, architecture is also always a mirror of the period. And from that pov it's not looking so great for beauty making it back any time soon.
[1] Yet functional because how can the happiness/joy of the perceiver of architecture be removed from what should count as functional, in architecture?
It is also interesting to read about the Amsterdamse School, an early 1900's style with "mass appeal" of which you can see many beautiful examples throughout the city, which were also only built thanks to strong support from the political sphere. https://architecture-history.org/schools/AMSTERDAM%20SCHOOL....
Surely raising capital by offering fractional ownership of the company that produced the god-like super-intelligence will be of no concern to the god-like super-intelligence or anyone else in short order . . .