Thank you so much for building something that was obviously good enough to sustain the community for 12 years <3
Turns out, the reason people wanted to use it was because the API was stable and every slight step towards improvement was met with... resistance. I stopped maintaining it because the burnout from seeing what it could be and not seeing any path to get there was too much for me.
I'm extremely grateful to burntsushi for writing the datetime library rust deserves. Chrono is stable and is maintained (though semi-deprecated) by an extremely reliable rustacean -- you don't need to leave, but you'll probably be happier if you do.
I really appreciate when a project calmly acknowledges that it's not the right tool for every use case, and is not trying to be. That's professional engineering.
Especially in the AI-age, the crucial skill will be what NOT to implement.
The alternative is to keep a relatively small standard library, and a good package manager, to make it easier for people to iterate on API design in third party packages. Major version changes in a single library can have much less impact accross the ecosystem than standard library packages. Eventually you may get a de facto standard, and sometimes they are then later incorporated into the standard library.
This model isn't without its downsides; it can sometimes mean fragmentation when standardization would help, sometimes these de facto standard libraries can become important enough that they run into the same ecosystem fragmentation issues on major releases. But it allows for more flexibility.
The latter model is what the Rust project decided on for everything outside of the most fundamental time types. Dates and times and timezones are notoriously hard to model in code, and easy to create API designs that are full of footguns, so this is a place where that extra freedom to iterate was more important than standardizing on an API early on.
That iteration has led to a number of different libraries, including time, chrono, and now jiff, allowing for exploring the space of API design to really get it right rather than being stuck with what was available early on.
Woah now, let's not get too controversial.
And the UN has repeatedly proven it does not have an effective enforcement mechanism.
Sometimes you do want to represent the concept of “this whole datatype OR nothing”. How do you suggest this is modelled? I think optionality is pretty fundamental. What’s wrong about classical null is that it ascribes optionality to all types regardless of whether they need it.
Being forced to represent optionality with a “special” default is terrible design IMO - see the recent SSH RCE on Musl, largely due to the poor choice of indicating “crypt failed”. Default values can be quite dangerous. Sometimes the domain you are modelling supports such a value, but having “default value” be the default (and not a choice) is a similar evil to default null.
Instead of hoping you remember to check for NULL, Rust forces you to make a decision on how to handle the possibility of failure. Sure, you can still pick a bad decision (unwrap) but that is at least a specific decision and you can’t forget to do it.
You can take it even further with things like the newtype pattern. If you have lots of numeric IDs (project ID, user ID, etc) you can give them distinct types so you don’t mix them up.
So Rust taught me I can program defensively against my own brain, which sometimes gets tired or lazy or distracted.
No, it's not. It's more like saying a writer should try their hand at different styles and formats of writing. So a novelist who periodically writes non-fiction essays or journalistic content or poetry (whether intended for publication or as a writing experiment), and that practice can in turn inform their writing as a novelist.
Besides, learning programming languages is much easier than learning natural languages. Most experienced programmers today know multiple languages, and most college grads in CS and Software Engineering graduate having used at least 5 or 6 languages if their school is halfway decent.
My favorite side quest that I definitely recommend. Physical parallelism is so fun to reason about.
I had the advantage of sitting next to a proper professional to bounces questions off of, so I was able to get an SPDIF decoder to work in about a week! It was a very interesting lesson in the costs of abstraction, and why you will never really find abstraction layers in hardware.
Just to counter any obvious moralizing, I am not saying using models like that is okay, but it's an entirely different kind of crime.
Gave me quite a funny feeling, tripping innocently over the empty shape that they've had to carve out. They wouldn't have spent the effort to do that if it wasn't pretty bad.
I deal with a lot of user-generated content, and I have been using OpenAI as a tool to help find offending content. Now I see that if CSAM ever makes it into one of my systems, OpenAI is going to send the authorities after me if their model sees it too, so that's a dealbreaker for that use case, and it was a big one!
It's also wild that they are trying to detect and report AI-generated CSAM. This does absolutely nothing to help protect actual children.
Just don't put it in MY face.
I don’t think this is something that can be solved by more surveillance though.
Most abuse (i think something like 80%? Not sure but it was a big number) happens within the family and between minors (minors abusing minors)
Changing how easy or hard it is to distribute or generate CSAM will help exactly nothing here even if that’s the goal.
If only.
The same "hard on crime" ratchet that happens to most every thing that's popular to punish and extremely unpopular to speak out against the tightening of said ratchet has happened here.
Another complicating factor is that it really blows to have humans look at stuff like that, so everyone would rather use automated classifiers; and if you assume good faith and share too much info with the affected parties, this will be exploited by the bad guys to gain insights into how the system works and how to circumvent it.
I don't think there's an elegant solution to this, other than "don't allow tech companies to get so big that they can afford not to care about individual customers".
Isn't CSAM only material based on abuse of real humans? How could computer models multiply CSAM by that definition?
https://www.nytimes.com/2022/08/21/technology/google-surveil...
Pornography is illegal to download or save in Singapore… I wonder how they were able to train a model without the data.
We asked for further information so we could report the uploader to the NCMEC and remove the content from our systems. They refused to give us any information about this incident, denied our appeal, and have so far, illegally refused our privacy requests for data about our account.
OpenAI does not care about fighting child sexual abuse, this is completely performative.
EDIT: These image uploads are up to 20+ years old, and are being served through Cloudflare with their CSAM scanning product enabled. We've received no notification of any problems from Cloudflare. Something on OpenAI's system thinks that an image we sent to them for classification is CSAM, and they aren't willing to elaborate or communicate about it at all. For reference, our website is https://ytmnd.com
I think this is the completely wrong takeaway.
Ober the last ~15 years I have probably used around ~10 web frameworks/libraries
_not a single one was even close to being complete or mostly correct, even just for standard HTTP ~1.1 handling_!!! (The routing, path, header handling. Not any of the complex stuff in HTTP2, Quick etc.)
the problem isn't using dependencies, that is the solution (not reinventing the wheel again).
And preferring reusing solutions which are more complete, practical testet and have more resources to maintain/improve basic functionality is (in most situations) not "anxious" development, but (should be) common sense.
The amount of hidden complexity in even somewhat basic stuff like HTTP is enourmous, rewriting everything all the time is wasting billions across the industry all the time and (in general) increases the chance for bugs.
The problem isn't "using dependencies" or small dependencies, but:
- availability (or lack) of high quality well maintained dependencies
- "no one" wanting to pay for software dependencies
- misguided tendencies for 'clever' & overengineered solution in one extreme (e.g. the stuff which lead to log4shell)
- misguided obsession with deceptive simplicity in artificial examples which lead to 1001 incomplete solutions because of the obsessiv simplicity doesn't stop at cutting out over-engineering but also cuts out fundamental parts
- lack of widely adopted, effective, tooling & infrastructure to trust dependencies.
- lack of tooling to more efficient reuse and compose working parts, the way we do programming seems just very fundamental broken.
- also a lot of legacy tech we are stuck with making many things unnecessary complicated
Also as a side note, number of packages != number of entities to trust. Weather you use 10 or 1 package from the same source, maintained and published in the same way often makes little to no difference for supply chain security.
The PHP2Shell bug existed more than 15 years ago, spanning releases as early as 5.x. There was no AI coding back then. So it's very hard to blame AI. There was a coding mistake, and AI found it in 2026. Probably a good thing before a malicious party does the same.
OSS definitely flourished way before 10-15 years ago. I'm not sure what the author suggests as "knowing the details". Should we all rebuild our own web scripting language to know "all the details": e.g. PHP? Probably not practical. A line has to be drawn on where your knowledge and expertise stop. THe line can be vague and consistently changing based on needs. But there will always be a line. We will always stand on shoulders of of our predecessors one way or another
I understand the author's trying to say the overall situation is unstable and trending negative but I am not sure the writing's logical flow is holding up very well.
A better way to put it is we've written a lot of code, and a lot of it had bugs hiding in plain sight. AI made the whole bill come due at once. And will this AI code explosion create another bill we won't know about until 20 years later? Who knows.
1. Consolidating internet-dependent projects to a few winning cloud vendors and a dozen winning SaaS providers, made all of those project have coincident failure points. Which is great for corporate accountability ("it's Amazon's issue, not ours!") but bad for end users ("shit, EVERYTHING is down today")
2. The movement to migrate fat local-first projects towards thin internet-dependent or server-hosted ones made almost all software share those same coincident failure paths. This was good for entrenching a subscription business model and brought some end user conveniences, but meant everything being down meant everything being down.
3. Anxious and lazy dependency-first development culture ("you could never write that as well as some github rando") meant projects now had an order of magnitude more contributors but only reviewed and owned code for a surface level, making all these projects extremely vulnerable to supply chain failures and subject to constant maintenance churn.
4. Employment and team tenure continually fell so that most projects no longer have experienced owners and instead just temporary custodians.
You can find the seeds of many those problems in the 1990's and earlier and you can even argue that some of them were inevitable (and even worthwhile) destinations for the industry as it evolved without an engineering guild to slow it down and insist on rigor or evidenced safety.
But anybody who's been around for a few decades can attest to how these risky patterns rapidly grew into norms over the past 20 years and shifted the industry from one where any given software had some idiosyncratic bugs that users could learn to work around to one where all software had lots of bugs, many of them correlated across apps, whete failures would wash through and shut down (or make vulnerable) whole industries for hours or days, with effecrively nobody in particular to hold to account for any of it.
Software has never been in a pristine "golden age", but the last 20 years built almost everything on one giant, shared house of cards that's now teetering on the brink of collapse. Better or wose, it's indisputably different than it was before.
As a result I get:
1) The enmity of my lead dev who prefers to do things more fast and loose.
2) Not promoted, due to the first reason and due to the fact that I don't have "impactful" metrics.
I'm not saying I'm perfect, there are definitely things I could have done better. But the fact still remains that the industry grossly undervalues maintenance and optimization that doesn't lead to easily measurable numbers that impact the bottom line.
That in turn is because it's hard to tell if SWEs are actually paying down critical debt or just wasting time. But that's the core problem, not that we're not good at it.
The good news is I can get AI agents to do this kind of work so I can now chase after metrics.
I find the arguments about decoherence and the sort of snowball-of-complexity more compelling than this timeline.
Maybe part of it is laziness but for sure part of is low morale and hopelessness.
AI is great at following patterns. AI is great at writing tests. AI is great at doing work that humans would consider drudgery and then applying "enough" judgement to it.
The problem is that most software teams and most engineers are very, very bad at the thing that AI needs to succeed: technical writing. Documenting the patterns, the rules, the heuristics, the SOP, the judgement; developers could barely be bothered to write comments before (which in the AI age are free, infrastructure-less memory that is always "line-of-sight" for the agent and can be used to enrich with keywords).
It's not that AI is the problem, it's that we humans are good at resolving these constraints in the real world at a human pace. AI makes that pace impossible for a human to keep up as a trailing activity. Rather teams need to invest more in technical writing and producing the corpus that AI relies on to learn the operational constraints and rules.
One consequence of that is that they scale more nicely for nesting: https://thatsmaths.com/wp-content/uploads/2016/07/squircles-... (right half of the image)
Of course, Apple’s squircles deviate from that by inserting straight line segments. What remains is that the corners still have a non-constant curvature.
Everyone loves to harp on about how terrible the Registry is, but that’s missing the forest for a tree.
The registry + GPO engine is a strongly typed and layered settings engine with:
Machine, Machine Policy, User, and User Policy layered on top of each other, consistently, through pretty much every native app!
More importantly, policy would apply or unapply dynamically and cleanly.
You could move a laptop to a different office and soon as it connected the wifi its set of configured printers would change, its outbound proxy could be reconfigured, etc.
More importantly: Microsoft engineers of the era put in the work of writing up thousands of settings in template files that shipped with Windows, with simple checkbox UI controls and paragraphs of descriptions in every language!
If you think your terrible “settings automation” tool is a viable alternative just ask yourself this question: does it “put the work in” or does its user have to?
Machine accounts have limited rights compared to regular users. AD provides fine grained distributed policy and cert managment. After decades of emperical refinement, AD is often superior to what the cloud offers including Entra. In environments with many PCs that require centrally managed policy, AD is still relavant and probably required.
I think it is also true that there are and always will be many companies that simply cannot expose these PCs to the cloud either because they're working with data that cannot be in the cloud like legal files, finanical documents, schematics or they cannot risk being tethered to the Internet. You know that little micro PC attached to the cart in every exam room in a doctors office? There is a 99% chance it is joined to AD.
> Human approval doesn’t scale to developers creating and deprecating testing
> databases, Kubernetes clusters, and CI runners at will. Contemporary AD
> implementations deal with this by deferring authority. An admin token (sometimes
> hardcoded…) automatically joins new identities. [...]
> Even if you built all of the complex automations to reflect every testing database
> and Kubernetes pod in AD with automated admin approval, it’s security theater. The
> unified directory would exist on paper, but lack authority. [...]
> Modern cloud infrastructure automatically asserts machine identity: [...] In
> Azure, you can configure a machine identity in Entra ID and let Azure’s Instance
> Metadata Service (IMDS) automatically issue OAuth tokens when machines request
> them.
Okay, so configuring machine identities manually in AD doesn't "scale" to the container age, fine. But configuring machine identities in EntraID does? How is auto-issueing identities in AD "security theater", but auto-issueing identities in EntraID the way to go?I believe what you're trying to say there, with "drop in replacement", is that jolt emulates common Java interop forms in Clojure JVM such that code written for Clojure JVM doesn't need to be ported to run on jolt. This a compelling feature which other dialects are tackling as well.
What I would mention as a key difference between jank and jolt, aside from the C++ story and native project management story, is that jolt is very much AI coded and jank is not. jolt doesn't disclose this on its website or Github readme, but it's something which clearly matters to a lot of people.
One thing to note is that the framing "Just 400 US Data Centers" makes it seem like 400 out of the gazillion data centers use this much electricity, but the original research seems to be that they found exactly 403 "Hyperscale Data Centers" and evaluated them.
The framing is technically correct (since there exist non-hyperscale data centers) but is clearly intended to be deceptive.
[1] https://hsph.harvard.edu/news/hyperscale-data-centers-energy...
Update (November 16): Answer!
Love the article! I work for Majestic
Great spot. the answer:
Some internal experimentation. Choose a domain, make it esoteric, do some experiments. However, looks like we forgot to add a filter to avoid reporting it. We will sort that…
—Steve Pitchford 0
Moral of the story: don’t test in production.
0 https://bsky.app/profile/did:plc:jhfnkrqilxtd3hfjkv7edswg/po...A well-known web security journalist, who writes in this mode, is Brian Krebs of https://krebsonsecurity.com/.
This type of ingestigative web-security writing has three audiences:
- the "general" audience of non-security-focused IT professionals, who are nevertheless interested in security topics from afar;
- the "professional" audience of other web-security professionals, who do the exact same kinds of investigations the author themselves does (whether or not they then share their results publicly);
- and the "oppositional" audience, i.e. the very malicious-actor groups the investigations target.
Each article thus serves three purposes:
1. to the "general" audience, the article provides an entertaining and informative journalistic narrative (and this encourages this group to spread the article, to increase the likelihood people in the other two groups will read it);
2. to the "professional" audience, the article serves as an "investigation log" that must provide enough information to reproduce the derived information (so that the results can be trusted and acted upon), and/or so that the investigation can be taken further without wasting time doing things that have already been done (because each piece like this, if it doesn't end in the investigation target being cleaned up, serves double-duty as an invitation for collaboration, either directly, or via encouraging independent "federated investigations");
3. to the "oppositional" audience, the article acts as propaganda for the power and reach of the "side" of the investigators, with the intent of producing chilling effects against others doing the same thing; and, more specifically to would-be collaborators with the target group, acts to scare them off from doing that, cutting off the target group's resources.
Note the three-way conflict-of-interest this introduces: security professionals need to know how and where you got your information; general audiences don't really care and just find such info to be needless detail; and your investigation targets should ideally know as little as possible about these tools, so that they don't start enacting countermeasures to make these tools useless against them. (But it's good that the opposition is aware that these tools exist and understand the chain of inputs and outputs that goes through them -- to show exactly the sloppiness that got them got, and so prove that web security people are not just bluffing about having such powerful tools while actually parallel-reconstructing info they get via e.g. double-agents.)
This conflict-of-interest is mostly resolved by just naming these tools while describing them as little as possible, and absolutely never linking to them.
Security professionals are expected to recognize the names and know how to get access to them (if they don't have it already); general audiences are expected to just gloss over these names as some kind of technobabble (an expectation you and I, being curious-by-nature people, failed); and the opposition is expected to recognize that these tools exist, but either struggle to find them (these tools are often referred to by nicknames that are also names of larger services in other domains, so just punching the tool's name into google often does nothing), or struggle to get access to them. (With a specific additional goal of minimizing the knowledge the opposition has of these tools so that they can't socially-engineer their way into convincing some bureaucratic employer that they "should" be given access to them. Which means never describing how these tools work, providing screenshots of them, etc.)
It could be an interesting way to observe if certain crawlers are obeying robots headers by analysing scraped content, or could just be some element on the page itself and part of turnstile.
If it was registered by the Cloudflare team, they might explain why.
is what he means. The phrasing is a bit of an idiom and isn't meant to be taken literally.
Anyways I kinda like the flippant(?) tone of the OP, wonder if there's a better term to describe this type of writing. Best I can think of is "halfway between image-board speak and casual speech".
https://www.antipope.org/charlie/blog-static/fiction/acceler...
Interestingly, it predicted the presence of mass AI robocall scams targeting the elderly almost a couple decades before it happened.
Majestic refers to https://majestic.com, "backlink analysis".
Tranco (from Gemini): "Tranco is a research-oriented, manipulation-hardened top one million website ranking list designed to fix reproducibility and methodological flaws in older popularity metrics like Alexa."
The other... No idea.
"The president of agalmic.holdings.root.184.97.AB5 is agalmic.holdings.root.184.97.201. The secretary is agalmic.holdings.root.184.D5, and the chair is agalmic.holdings.root.184.E8.FF. All the shares are owned by those companies in equal measure, and I can tell you that their regulations are written in Python. Have a nice day, now!"
For my team projects, I still need my teammates to review code, so that bottlenecks how quickly we can burn down the backlog. If I didn't need code review I would have the backlog burned down in weeks, but I'd create a bunch of outages in the process (we do catch things in code review!)
Also, some of my solo projects are gamedev-related, and gamedev is also bottlenecked on actually playing your game to see if its fun :) but it is great to get out of a 2 hour playtest session with notes and have a new build in like 15 minutes with the feedback incorporated.
This is mostly a problem of agents not having the autonomy of managing an entire project and having the ability to make tradeoffs to increase simplicity or maintainability, but when given that autonomy no AI model I've used uses that privilege to write simpler, more maintainable code.
I'm unsure about the energy efficiency of Tenstorrent cards, but they are at the moment much more cost effective than Nvidia. But software support is far behind, I suspect if it wasn't the cost would be more equal.
That said, Nvidia is in a strong position.
Even in a slowdown, they probably would do ok.
That's bad math. If you had 100k invested, and you're up 100%, then you were at 200k. A 20% drop leaves you with 160k, and you're only up 60%.
So much better than the app and nobody can take away or change the content :)
Overall it's frustrating and feels like it's intentionally designed to trick you into buying things you don't want.
God I miss being able to walk into a store to find a quality item that I need. Nowadays it's really only Microcenter where you can decide "I need this thing" and come home with that thing.
At some point, we lost the plot on online shopping. Sometimes physically going to the store to hold an item in your hand and bring it home within the hour is far more convenient than comparing 4-8 items online and waiting several days for it to arrive.
We really lost a lot when almost all shopping went online.
Since I stopped using Amazon I realised most websites now have a super easy purchase flow with Apple Pay or Stripe and the convenience factor just isn't a big difference any more. I also most of the time pay less than Amazon.
Black Friday predates Amazon by decades, by the way.
The only method I’ve found to work with Aliexpress is to add each item from each vendor to cart and get all the way to checkout to see what the real price will be. It’s a lot of work, and then I discover they all end up about the same price with discounts factored in and feel stupid for getting tricked into playing their game.
Not just jacking up prices and having fake sale prices (they do that year round) - but a specific pattern to throw off naive URL or SKU based price trackers:
They delist old models, and add them back in, sometimes with different model names, but the same specs. The week before the big sale, there were hundreds of delistings, and the average price rose nearly $80 - so that the $50 off becomes an increase.
https://minipcs.zip/aliexpress-11-11-pricing-tricks/
I've been tracking Mini PC prices on Amazon for months now and am excited to see if the same pattern repeats for Black Friday. (There's other weird patterns on Amazon, such as listing the same PC 10 times with different prices, all from the same seller, I guess to profit from consumers who don't research much?)
Most online retailers (Amazon included) still give free shipping above a certain threshold ($35 per order or whatever) so for things I need, I wait until I reach that threshold to place the order, and often purchase at other retailers or directly from the brand.
Also, everyone does "Black Friday" sales nowadays, in many cases matching the same window as Amazon.
https://en.wikipedia.org/wiki/Ron_Johnson_(businessman)#JCPe...
Also most things you can find on AliExpress for cheaper. Don't do edibles, beauty products etc though. I limit my purchases to tools and electronics
20% Off! 8€ instead of 10€ (lowest price in the last 30 days: 5€)
and you know it's a junk sale.
Walk into any store (JC Penny etc) half the items are on sale. Even groceries - go to Safeway, there is a newspaper style coupons - multiple pages. This whole thing is just bizarre and tiring.
One of my resolutions this year is to always search for independent sellers (online or brick and mortar) to compare before buying products on Amazon. In nearly all cases I have found a webshop with a better selection and a lower price, even after paying shipping.
It turns out that people like to do some research, look at a lot of images, read user reviews and more before making a purchasing decision. The subset of people that treat most everyday shopping as a completely automatable commodity is much lower than the tech world thinks.
This is a stupid meme that became a truism that really needs to die.
Stop mystifying basic network services.
All you're doing is scaring people off from running their own independent services leading to further consolidation and monopolization of the web
I also think that Maj. Ramsbottom is probably walking into a world of pain in order to do the right thing, particularly for those who serve under him.
> the extension internal UUID which is random per-installation
... and it's possible for a loaded page to figure it out. It sure sounds like this gives us a stable unique tracking ID ... ?
https://developer.mozilla.org/en-US/docs/Glossary/Top_layer
As a fun fact the Devtools also have a separate "Overlay Page" to draw on top of the web page for specifically this reason. The date picker on some platforms similarly uses an overlay page. This machinery was just never exposed to extensions.
You've been required to use a browser based ad-blocked on YouTube for over 5 years now.
If this sounds appealing to you, you may be happy to know that XUL lives on as Unified XUL Platform (UXP) and is still supported by currently maintained browsers.
UXP: https://repo.palemoon.org/MoonchildProductions/UXP
Pale Moon: https://www.palemoon.org/
Basilisk: https://www.basilisk-browser.org/faq.html
SeaMonkey: https://www.seamonkey-project.org/
uBlock Origin: 1.75.4
uBO Lite: 2026.1110.1816
Oh great. I recently learned[0] that some Firefox extensions can be tricked into leaking their install-unique UUID, which can then be used as a tracking identifier. Today I learn that any site can get my unique uBO UUID by hooking fetch('/adsbygoogle.js').
That is... an ironic disaster.
What you need to watch out for are the people or companies who latch on to proof checkers and formal verification as some sort of shortcut to correctness, intending to use it as a signal of seriousness rather than a serious tool in its own right.
It's refreshing to see intellectual honesty about a tool's limitations, from someone still advocating for its strengths.
It is worth noting that of the 290 OpenAI theorems which come with Lean code, around 38 of them have the property that the Lean code does not even state, let alone prove, the main theorem of the paper.
So even if LEAN is 100% bug-free, its still on the programmer to correctly express the logic and proofs, or they could be proving completely unrelated facts.
That it's the equivalent of saying "It's full moon tonight because I wrote this program".
And he states that you have to check that:
1. The statement the author is making corresponds to the statement in the lean program.
2. The lean program actually compiles.
3. There is no shenanigans hidden in the lean program to disguise a trivial proof as the one stated.
4. The program does not exploit bugs in lean.
Sure, the code works, but does it work at doing what you want?
Sure, the proof is valid, but does it prove what you set out to prove?
the author of the paper avoids giving any weight to two big features:
a) all male spaces (not true of modern cafes) b) almost exclusively white/european
it is pointed out that the coffeehouse is a non-European invention - and some of the original proprietors were non-European - he even examines why coffeehouses in the Arab world didn't produce scientific progress - his answer is that they were suppressed by the powers that be
the British pub is the closest thing we have today - but they increasingly skew old in age - and are too much IMO "places to drink alcohol" "places to watch football" vs what the coffeehouses represented - they are also relatively expensive
Ordinarily, if you want to adjust the parameters of a CAD part, say, a bracket, to meet a constraint like a particular mass or center of mass, the approach is to change every parameter a little, rebuild the geometry to see what physical properties changes, and repeat until you get the result you wanted.
This is of course a very expensive operation because you have to iterate through every parameter on a part every single time for any optimization pass.
So VCAD's approach is to essentially record how each of the mesh verticies connect to each other after the BRep tesselates, then calculate what part of the mesh are actually affected when a parameter change, and then use that to guide their optimizer to get the physical result they wanted, so they only need to change the relevant parameters they found that affects properties instead of doing the computational equivalent of banging your head against the keyboard and hoping something works.
They could have been a niche gaming company, and four years ago their stock price was...$12 ...yeah 12 dollars, that is a bubble for you, and can go back to $12 as quickly.
He got hit on the head with AI....And essentially, got the lottery ticket of the mechanical sympathy of GPUs, having the ideal architecture for the core algorithm of a technological revolution, that only happens every 50 years.
A computer science advance can as well quickly, make their GPUs obsolete. The Jev model being a possible example, even if yes, you still need the GPUs for Training.
It was not long that for their talk of AI, they were hanging up the phone on Geoffrey Hinton, when he was begging for a donation of one GPU: https://youtu.be/n4IQOBka8bc?t=1352
You need the merit to be at right place and right time, but the analysis of what NVIDIA needs seems to forget the most important: The need their luck not to turn...
If you know CUDA in depth, you will know there is not much there, its not a deep layer, and I am surprised ZLUDA or SCALE are not taking more adoption.
Steve Jobs was completely unsentimental about obsoleting Apple's lineup with new products but that's because Steve Jobs's strategy was to be a ruthless fast follower. Just wait for the technology to catch up with a burgeoning product category and then come out with a product with way more polish and sensible UX than everyone else. That's great for consumer electronics but its not so great for Nvidia. There's a lot less value in "taste" and "polish" when it comes to a device that's going to sit in a data center for its whole life.
The Startup Owner's Manual and The Four Steps to the Epiphany are fine books ... Certainly somewhere in my top 40 startup books you should take the time to understand. Like Geoffrey Moore, Ries & Trout, Clayton Christensen, somewhere in that group.
You'll be in for a commitment though.
A lot of popular "must read" business books are just folksy fluff ... blank is the opposite.
Also: AirPods are really good, again leading the entire market.
Sadly they have not done nearly as well with the Mac and iOS software, which disappoints regularly.
From momentum. Apple is a lifestyle brand. There is no reason for the normal consumer to want the latest iPhone over the latest two iterations. And yet, a lot of the new owners got it from Model -1 tradeins, because it has the one thing the other ones dont" identifiable camera configuration when seen from the back. It signals "I own this".
95% of software products hold their seat because they got the seat in the first place years ago, and not because they're necessarily the best.
Everything after that happens in the light of the absolute knockout success of the iMac, which got a staggering amount of good press and bought them so much earned media because it was such a product-placement-friendly device; as was the original iBook.
(My low key odd belief is that one of the biggest positives of that era for Apple was the Mac version of Outlook Express. I think we forget just how much easier the iMac made going onto the internet)
[0] you could argue that Apple was, in a way, too culturally important to fail at that moment; Microsoft invested (with non-voting stock) as a way to ameliorate some of their anti-trust circumstances, because the weakened state of Apple was perceived by the press and the world of politics to be partly a consequence of their behaviour.
Apple did not invent or even popularize smart watches, fingerprint readers, Bluetooth earbuds, smart home hubs, VR/AR headsets, self driving, mobile payments, foldable phones, etc.
About 2/3rd of the Apple versions of these aren’t even the most popular in their category.
Also, sleeping with an executive of a customer reflects very poor judgment, even more so when the customer can be somewhat credibly be accused of export violations. Then, on top of that, he sleeps with a subordinate. A day later, it seems there might be a second employee. This isn't like he's sleeping with two or three random women that aren't his wife, it's employees and customers. It would be hard to find worse choices. He's tarnished his own brand. The Nvidia board did the right thing. (Also, if you have a bus factor of 1, your first priority is fixing that.)
I don’t like Nvidia, but I hate private chips far more.
Maybe in the long term. The biggest threat in the short term is Nvidia's exposure to the data center bubble, which was already showing signs of popping before Jensen's departure. Given Nvidia's role as a key investment bank in the AI trade, making and brokering circular financing deals around data centers, his departure could hasten the data center collapse if the company starts backing away from its key financing role.
1. Physical infrastructure - power, cooling, network fabric, chip fabrication
2. Chips
3. Low level sw - CUDA/ROCm/Metal/XLA etc
4. Kernels - cuBLAS, cuDNN, FlashAttention, FlashInfer, Triton, llama.cpp
5. Kernel gen/optimization - MLIR, TVM, TensorRT
6. ML frameworks pytorch, jax, tensorslow
7. Models - Llama, Qwen, Claude, GPT, Gemini
8. More tuning - vLLM, SGLang, TensorRT-LLM, DeepSpeed, Megatron-LM
9. API/SDK/Platform
10. Apps/Agents/Pipelines/Orchestration etc
The biggest moment was partnering with Motorola as phones started becoming more capable and flash memory got cheaper to bring music to phones, which was then the prelude to the iPhone.
The iPhone killed the iPod. Most companies would have never created the product, because it would have killed their golden goose, but Jobs did.
Just like he went all in on the iPad which had the potential to kill the cheap student Mac.
If Nvidia's next CEO is smart, they'd buy up a bunch of these companies that "burn" weights to silicon, make the process better, and then sell that as a service. Let it cannibalize the sales of their products. It's better for them to do it to themselves than their competitors.
Surely Nvidia won’t stand still and will continue to compete, but the margins will collapse along with volumes.
AMD will probably suffer less, relatively speaking. Their failure to invest in software finally seems to be paying off now that LLMs can pick up their slack. But the growth currently priced in to their stock will never materialize and they’ll be seen as arriving just in time for the lights to go out.
Neither option is perfect.
A third option would be updating the benchmark every 6-12 months, but that has its challenges as well as you need to make sure that the benchmark covers things the model hasn't seen or is harder for the models to game.
This is wrong.
To make a vacuum analogy, one brand's Low setting could be another's Medium setting.
To compare you either give them the same resources ( Watts or $ ) and see how much they can do, or you give them a task and then check how many resources they required to finish it.
The label on the knob is irrelevant, except that either model might have a setting higher than the other which is sometimes useful if the resources aren't a constraint or lower when they are.
There was also a bug in which the reasoning tokens were “encrypted” for opus but you could ask Haiku to regurgitate them and it would.
Anthropic have been doing anything and everything to prevent distillation “usage” and it’s caused me at least once to cancel my subscription because it’s hard to trust a company that secretly alters its responses based on something I can’t see and never get feedback on.
They have altered their benchmark multiple times in the last 4 weeks (v5 in https://artificialanalysis.ai/articles/artificial-analysis-i...) and the last 5.0.1 change https://artificialanalysis.ai/methodology/intelligence-bench... pushed up Anthropic's models a bit, you can see that in v5.0.0 at the first link, Opus 5.5 has a lower score than in it gets in v5.0.1, and v5.0.1 was released just before Fable dropped.
Some part of me believes that the benchmark mix was put together to favor Anthropic's models.
So, with new models coming out on a weekly basis, I think trying to figure out a future trajectory is just going to be a wild guess. For all we know, Anthropic and OpenAI models may also be plateauing in the near future; or there may be a breakthrough and they may improve beyond our wildest expectations. I honestly don't know.
That said, I suspect available compute and model sizes that major Western vs. Chinese labs can handle will still keep the gap open at least in the short term.
On the other hand, I am also not sure how much the gap will matter past a certain point. I am specifically wondering at what point people will become interested in smaller models fine-tuned for specific application areas (e.g. cheminformatics) rather than massive generalist models.
Scroll down to the cost per task pareto front chart which includes more difficult levels by default: https://artificialanalysis.ai/models/qwen4-max#intelligence-...
Qwen4-Max is not in a good spot on that chart. Astra xhigh and Fable 5.2 medium had a similar cost per task with higher intelligence ratings.
For comparison, here's one I got from Qwen3.8 Max (at default reasoning effort): https://simonwillison.net/2026/Jul/20/afraid-of-chinese-mode...
This inspired me to re-run pelicans through 3.8 Max - but OpenRouter has qwen3.8-max-prime and qwen3.8-max-0902 - prime is twice the price and faster - so I went with the 0902 one.
Turns out Qwen4-Max has a none setting, but 3.8 Max didn't allow disabling reasoning - you had to pick from these five levels instead: minimal, low, medium, high, xhigh.
Here are those 3.8 Max 0902 pelicans: https://tools.simonwillison.net/markdown-svg-renderer?url=ht... - they're mostly not as good as 4, but I prefer the 3.8 xhigh one.
[1] Especially Terminal-Bench-Science 0.1 stands out. Example problems:
- https://hub.harborframework.com/tasks/terminal-bench-science...
- https://hub.harborframework.com/tasks/terminal-bench-science...
Probably very relevant to people working in these fields, but it is honestly not surprising if a lab with limited compute will not train for these technologies, and it says pretty much nothing about more common use cases. I also suspect that this benchmark measures knowledge retrieval more than raw reasoning ability. This is where you really need a breakdown of capabilities and not an increasingly meaningless aggregate number.
By comparison opus 5.5 (max) is $10.01.
If they are not testing Chinese models with less than max thinking, then we should only compare them to other models also at max thinking.
Edit: In addition to reshuffling the weighting of individual benchmarks, they also do things like test Anthropic and OpenAI in their native harnesses, while not doing the same for Qwen/Deepseek/etc.
And it’s a shame because the only reason it scores so low is that it’s slooooow. (Like 30 tokens per second vs opus near 80.) If you exclude voting, Qwen 3.8 does well.
(We allow early voting contrary to Arena.ai because I think it makes sense for agents, where it doesn’t for chat. This means our data takes speed into account.)
Also, it would be ideal if they served this with third parties. Part of the slowness with 3.8 is just not a lot of providers. Their open source version is much faster, but we couldn’t include it because it doesn’t handle images (another own goal by Alibaba).
Looking forward to the "Flash" and 27B versions.
Humans can also be trained in tasks like these and we improve a lot over the course of training, but my ADD banana brain for one probably won't ever get past 50-60% of what the chimp can do.
Among parrots, everyone knows African Greys but cockatoos are also highly intelligent, with goffin's cockatoos the standouts in terms of cleverness.
Octopuses have a much more distributed system than mammals, suggesting a quite different approach is possible.
I hope folks are re-evaluating that now.