1. Commercial devices must use offline only protocols, Zigbee or Zwave. If it's WiFi only I will not buy it unless the firmware is open source or easily replaced with an open source, local only alternative.
2. Non-sensor devices must fully function without their IoT features. I must be able to control my thermostat, fridge, stove, etc without any external interface.
Never use a WiFi IoT device that isn't open source. Even if you can control it locally it could be a spy device that's exfiltrating your data. Yes, you could setup VLANs or a separate network but it's not worth it. Vote with your wallet.
Reminds me of a possible apocryphal anecdote I heard about Alan Turing. He loved going on long bike rides, but one day his bike developed a distressing habit of slipping the chain. After the third time it happened, he realized it was due to a broken tooth on both the front and the back gear. He went back home and calculated how many revolutions of the pedals it would take for both broken teeth to line up when they were at the critical part of the rotation, and he built a little mechanical counter that would click on each link in the bike chain, and at the calculated time, a little arm would reach down and clasp the chain, holding it in place long enough for the teeth to re-engage.
He showed the device to his friend, who said "Why didn't you just buy new gears? You could have gotten both for a dime." And Alan replied "Oh, that's clever, I didn't think of that."
I only use "Local Polling" or better, and highly prefer "Local Push".
In a moment of weakness, I bought a D-Link wireless (wifi) water detector for the sump pit (it was on clearance for $30); it relied on a D-Link Home cloud service (free) that was closed down a few years ago.
Not wanting to toss it, I repurposed a spare Chromebook 14 that I was going to recycle and turned it into a Alpine monitor for the water detector, which I'm powering from the Chromebook's USB port (if power is lost in a rainstorm, I need to know about it before the sump pit fills up).
An immediate issue was that the D-Link firmware tried to "phone home" every 24 hours and would drop from the local network, since the server was kaput, and I had no clue what response it was expecting from the server to keep it alive. Solution; "reboot" the unit by interrupting USB power (with a Pico and MOSFET) every 23 hours. There's no question that home automation should work without the internet.
1. buy the Zigbee version
2. boom, local only smart home is done
I guess if there's a device that's actually wi-fi only then I'll have to go through your flow but so far there's been no need.
- buy the cheapest hardware that gets the job done
- ignore negative reviews that are software related
- upon receipt of devices, ask claude for local-first cfw
- have fun for a few hours acting as a meat proxy
- boom, cheap local-only smart home is done
Use a SBC or nas for home assistant. This process worked for me for switches, a thermostat, an air quality monitor, a thermometer, and a window fan.
Would’ve paid ~2.5x for hardware that had the appropriate local-first capability out of the box. Instead this all fit within an existing $20 subscription to an inference provider. DS 4.1 running locally was on track too just more slowly.
I managed to get to:
The cloud is the weak link
- No internet, no problem. ...
---
If you are writing for other machines to ingest it, whatever. But if you are writing for people it for me at least signals that you don't really care about the message that you want to get across, and are fine with a watered down LLM average.
1. numbers as the gateway to quantitative thinking
2. algebra as the gateway to abstraction
3. geometry as a way to calculate beyond what we can see
4. probability as a tool to navigate uncertainty with rigorous thinking
5. analysis as a means to tame the very large or the very small
6. dynamics as the mathematics of change
Terence Tao explains 6 essential mathematical concepts: https://www.youtube.com/watch?v=OOMx2BHHWtE
The thing is the parents normalizing this kind of intrusion of privacy think they are in the right while creating a very dangerous world. Even things like going through your kids rooms (yes even when they live at home) is a huge privacy violation.
No wonder the "but think of the kids" argument is so effective, we all grew up being propagandized by or own parents and environment.
They hated us for it. Other parents called us crazy. “They should learn how to handle social media!” And now their kids are unmanageable and ours thank us for shielding them.
So parents keep their children indoors for fear of being arrested from them crossing the street on their own. Politicians are unable to renege on "child safety" laws because the populace will believe they are harming children and vote them out.
But these things work the way they do because of the social consequences embedded in our culture, caused by the perceived harm to children from inaction. People want to avoid accusations of being complicit in harming children by supporting politician X or failing to do Y, as that has legitimate societal consequences when taken far enough, so there little incentive to diverge from this narrative.
It's actually sad for kids. In childhood being able to stay out a bit longer, solve sudden problems without calling mom, or just sit alone taught you how to cope with the world. But if someone monitors a child all the way to their college dorm door, it's unclear how they're supposed to develop personal autonomy at all
I don't agree. I activated most of the protections. If a site breaks, then I don't get to see it. That's OK. No big loss. Users should fight their FOMO.
Steve and Tom also individually have some of the best blogs out on the web.
I wish them all the best with whatever comes next. I'm pretty sure they'll all do great things.
Nothing to to do with the quality of Val Town, but making a business in this space is really, really hard. You can't start the business without raising millions (Eve raised $2M, Darklang raised $3.5M, Val Town raised $7M), but it's hard to scale out a proof of concept in the time investors want to start seeing an ROI. It seems to me about 2-3 years is all you get for seed runway, when realistically projects like this should take at minimum 5 years incubating. Darklang extended their runway to ~7-8 years by firing all the devs and keeping just the founder on life support. Ultimately even that wasn't enough.
But maybe with AI this could work, I dunno.
What's required to make it all work is a fundamental leap in programming interface design, but the business side doesn't look any different from a typical SaaS. So unless you hit the ball out of the park you're entering the market with an experimental product, then you have to deal with product-market fit even if you have the right product, so you really only have one shot.
It's very hard to go back to investors and ask for another round of funding when there's no discernable market and the product isn't gaining the traction investors would like to see.
Anyway, kudos to Steve for fighting the good fight! What would really be great is to read an in-depth post mortem on the project. Looking forward to it if you have time!
I'm definitely looking forward to seeing the code in February when they open source it.
>I did my PhD in computer science at Stanford, where I designed long-context architectures for sequence modeling and memory. Before that, I studied math & physics at Cornell. Most recently, I was a research scientist at Cartesia, working on in-context tokenization and novel architectures (HNets).
Not the MSL Alex Wang
WireGuard by itself is not a stealth VPN, but the things that make a VPN "stealthy" are all things you'd layer on top of a WireGuard design, not thread through the middle of it. A datagram-based protocol can run under any obscured transport.
We did this at Fly.io when we discovered that, for one reason or another, 5% of our users simply couldn't reliably use UDP. It was a one-day project to run it under WebSockets instead, and that was before coding agents.
Then instantly gives you a shell script that it pipes to `sudo bash`. The irony isn't lost here.
I expect that Anthropic will have success in some of their side hustles if they're reinforced by the model ecosystem and double down on them, but you should also expect that at some point in the future we will talk about the "Anthropic Graveyard".
Dots in particular display some insidious methodology, they've hidden the "programmable" portions of an agent. They want to make it 'magic' (all your work goes through a single tool) but... that's pretty much exactly what businesses DON'T want. We need to see what it's doing, accessing, working with, underlying data, etc... It's an explicit business requirement for us.
The more they can 'hide' what the backend does, the harder it is to get away from the product.
Buuuuut, I don't think it's going to work. I can still export it somehow if I want to, and it would probably take 10-15 minutes... The whole point of these agents is that they're 'plug and play' with whatever technology you want.
And if they don't build them this way, someone will whip up an open source version that does it cheaper and more transparently in a couple of weeks.
Software being free now hits both of these companies exactly where they want protection.
I read this as: We'll build anything and everything into Claude eventually.
How does this work long-term? Is the new reality going to be picking an AI provider for all of your SaSS tooling, similar to Cloud providers for infra? Hopefully competition thrives.
I still choose .NET as the go-to whenever I can. On Windows or Linux.
I think the baggage view was valid for a while - but now we have a new generation of developers who only know about .Net Core and may not associate the Framework years in quite the same way.
I got so worked up by it I built my own TyepScript compiler with my clankers in C# to prove that it can be done, my revival of the Common Language Runtime concept, project Copeland:
https://github.com/yuechen-li-dev/copeland
Currently, Copeland compiles a dialect of TypeScript to JavaScript, C#, WASM via Blazor WebAssembly, and NativeAOT, runs in csproj, and uses xUnit.NET for testing framework.
(The other reason is that Microsoft still refuses to do the obvious thing in .NET 11 and take the best part of Rust and let enums be `match`ed, despite the fact that both TypeScript and C# have weak enums historically, so I added payload enums to my TypeScript dialect because Microsoft won't.)
> Being a Microsoft product is such a heavy baggage
Which I always found ironic because TypeScript (from the same guy that created C# even!), VS Code are both Microsoft. Microsoft also makes big contributions and investments in Postgres and Python as well.It's a good language though. The latest trend of hating OOP is unjustified/overreaction. I do write Rust too and I am not really convinced that traits are better than shallow inheritance + interface-oriented classes.
Some folks out there still think .NET is Windows only. Some folks out there haven't looked at C# since the .NET Framework days. Some folks out there are still salty about some of Microsoft's decision around OSS (debugger, hot reload fiasco, etc.)
But day to day working with C# and .NET for backends? It's really fantastic. AI seems to like it a lot and a lot of tools for agents to go HAM. Roslyn analyzer, source generators, workable hot reload (for a compiled language), tons of really good docs from Microsoft, generally low variance and high consistency from codebase to codebase (e.g. encountering Drizzle vs Prisma vs TypeORM vs Kysely vs ... in Node land).
My previous team switched from TS to C# and F#[0][1] and it was a blessing; a lot of the platform ops have been simplified. It's just one very solid tool chain and packages are generally stable (a lot less effort on maintaining platform config/tooling and more focus on just building). First party packages get consistent security updates and C# doesn't have nearly the same level of exposure to transitive packages that one encounters with JS/TS ecosystem.
C# feels like a really solid choice in the age of AI coding because of it's general stability, strong static analyzer, insistence on types, and really good tooling. Not right for all use cases (e.g. CLI, serverless functions), but fantastic for backend APIs.
If you're looking to do AI with/on C#, I wrote about some of my experience with making codebases successful for AI[2].
[0] https://engineering.usemotion.com/moving-off-of-typescript-e...
[1] https://typescript-is-like-csharp.chrlschn.dev/
[2] https://chrlschn.dev/blog/2026/08/the-unexpected-ai-stack-cs...
Be aware
(decimal)0.1
// .NET 10: 0.1
// .NET 11: 0.1000000000000000055511151231Once burned, twice shy.
I wound up going to college after working retail at Disney, then after college, my first job was supposed to be C# but they didn't have the project approved, so I got my first exposure to Python - which I loved, then JavaScript - which I hated but have a love-hate with to this day, and finally C#, but it wasn't .NET 3.5 no, they started using .NET Core, which was... something else, lots of teeth pulling during the original .NET Core days.
So many wonderful projects made with .NET my only complaint is I wish Microsoft would make a cross-platform GUI stack they'd actually stick to, and make it work on Linux too. ;)
I know there's alternatives but nothing beats telling your coworkers "lets just use the built-in" when choosing a GUI stack.
My experience with Go is a while back, before it got generics. While I liked the overall idea of a language with little syntax and a focus on simplicity, I didn't really like Go in the end. It was a weird mix of too low-level and high level code. It exposed low level details that I don't need to think about in C#, and it is missing convenient abstractions like LINQ.
Personally, I like async/await in C# more than Go concurrency. I know this is heretical here, but I don't care about function coloring or any of the abstract advantages of other concurrency approaches. Async/Await in C# works for me (in non-legacy code) and is easy to understand.
Also Go is at the stage of C# .NET 2.0, and still plenty of lacking language features.
I really think the creator of C# made a disservice to the .NET community by picking Go for the Typescript rewrite.
I have also found c# async to be lighter than go's concurrency story, and dotnet11 is going to make that much more clear.
Now: I write software for fun, not for money.
The strange ending does also perfectly align with the dreams of a singularity lots of people here seem to have. We are so far from that.
It was of course written with Claude. Have fun spotting the tells but also there’s a weird flatness to the writing, like the uncanny valley for prose.
Yes, I view most 'new' TLDs as scams. They lure you in with a low price, a rather pricey renewal and then price hike after price hike down the line. They also charge a premium for most dictionary words right from the start.
Unfortunately I am stuck in commitments on a few of these domains and just had to eat the cost increases, the rest I moved to .com/.net.
.consulting was the worst offender.
The question before the Court was not whether a domain name is a form of property. Both parties agreed that a domain name is a form of property; that wasn't in dispute. The Court did affirm their shared position in black letter law, though.
Rather, the question before the Court was whether, as intangible property, a tort of conversion claim relating to the domain name could be made under California law. Previously, cases disagreed as to whether it could be recognized. The Court found that it could be.
And to show off another fun character, for languages like Spanish that start a question with an inverted question mark, there's also ⸘, the inverted interrobang.
Though I find the name some people call it, gnaborretni (interrobang backwards), much sillier.
https://en.wikipedia.org/wiki/Kremen_v._Cohen
but it's tricky because the domain name registries and registrars have a kind of thicket of contracts nowadays that subordinates registrants' property interests to other rules and considerations. So I don't think a court would think that a modern-day domain name registrant has property rights in a domain name that supersede all of those contractual terms.
Edit: my recollection is that Kremen v. Cohen interpreted the legal defaults that applied because Kremen registered the domain name in an era when Network Solutions didn't require registrants to agree to a contract that limited Network Solutions's liability and gave it various explicit rights and powers.
I'm not a huge ICANN expert, but I think ICANN broadly thinks the status quo is preferable to one in which registrants had stronger, less equivocal property interests in names.
I did notice that my domain kept working past the supposed shutdown date, but I let the domain registration expire.
I'm so mad now! I assume there is no way to get my domain/email forwarding back :(
I suppose I could try to ask the registrar, but I highly doubt they can do anything...
(Edit: I'm not actually mad, but it was annoying to have to move accounts from my first@last.name email to another one)
Many registrars don't support the unique 3rd level domain setup for .name, so it can be hard to find registrars that properly support it.
There's also an associate email redirect service, which is unique to .name and probably costs money to run and maintain.
It's probably easier to support, and more valuable to sell second level .name domains.
Pretty terrible that they were considering just dropping tens of thousands of customer registrations however.
This was what I figured SHOULD happen when they did this. Allow existing customers to keep renewing, but nobody new allowed, if a tld frees up over time, reclaim it completely and sell it off.
What they were proposing should really be illegal...
Resistance with a strong case and competent lawyers is not futile.
> This post documents how I’m using various secure enclaves (YubiKey, TPM, Apple Secure Enclave) to hold/encrypt SSH key material and achieve the goal of not having any unprotected key material living on-disk.
Could you help me out and state the goal of your practices as simply as this? I can guess, but I'm really not sure, especially when you get into nftables checks for window sizes.
- Put the public authorized keys in /etc/ssh/keys/%u immutable, monitored for checksum changes.
- Have CIDR restrictions in said public keys.
from="x.x.x.x/n,x.x.x.x/n" ssh-ed25519 AAAA[snip] Text-Identifier
- Have NFTables checks for MSS, TCP SYN TTL, Window sizes, CIDR blocks permitted on the public ssh instances (throw away nodes) and behind a VPN on more important roles. If I have a need to keep netfilter disabled then I put sshd behind L4 haproxy with ACL's.Of all the things I'm looking forward to on Zen6, an improved memory controller is at the top of the list.
Edit: and yes, others may say the paltry few PCIE 5.0 lanes is the worst thing about AM5, but that doesn't affect my usage much, and unfortunately I think that's a limitation of AM5 itself rather than the I/O chiplet.
> Being a Microsoft product is such a heavy baggage
Which I always found ironic because TypeScript (from the same guy that created C# even!), VS Code are both Microsoft. Microsoft also makes big contributions and investments in Postgres and Python as well.It's a good language though. The latest trend of hating OOP is unjustified/overreaction. I do write Rust too and I am not really convinced that traits are better than shallow inheritance + interface-oriented classes.
Some folks out there still think .NET is Windows only. Some folks out there haven't looked at C# since the .NET Framework days. Some folks out there are still salty about some of Microsoft's decision around OSS (debugger, hot reload fiasco, etc.)
But day to day working with C# and .NET for backends? It's really fantastic. AI seems to like it a lot and a lot of tools for agents to go HAM. Roslyn analyzer, source generators, workable hot reload (for a compiled language), tons of really good docs from Microsoft, generally low variance and high consistency from codebase to codebase (e.g. encountering Drizzle vs Prisma vs TypeORM vs Kysely vs ... in Node land).
My previous team switched from TS to C# and F#[0][1] and it was a blessing; a lot of the platform ops have been simplified. It's just one very solid tool chain and packages are generally stable (a lot less effort on maintaining platform config/tooling and more focus on just building). First party packages get consistent security updates and C# doesn't have nearly the same level of exposure to transitive packages that one encounters with JS/TS ecosystem.
C# feels like a really solid choice in the age of AI coding because of it's general stability, strong static analyzer, insistence on types, and really good tooling. Not right for all use cases (e.g. CLI, serverless functions), but fantastic for backend APIs.
If you're looking to do AI with/on C#, I wrote about some of my experience with making codebases successful for AI[2].
[0] https://engineering.usemotion.com/moving-off-of-typescript-e...
[1] https://typescript-is-like-csharp.chrlschn.dev/
[2] https://chrlschn.dev/blog/2026/08/the-unexpected-ai-stack-cs...
Be aware
(decimal)0.1
// .NET 10: 0.1
// .NET 11: 0.1000000000000000055511151231Once burned, twice shy.
I wound up going to college after working retail at Disney, then after college, my first job was supposed to be C# but they didn't have the project approved, so I got my first exposure to Python - which I loved, then JavaScript - which I hated but have a love-hate with to this day, and finally C#, but it wasn't .NET 3.5 no, they started using .NET Core, which was... something else, lots of teeth pulling during the original .NET Core days.
So many wonderful projects made with .NET my only complaint is I wish Microsoft would make a cross-platform GUI stack they'd actually stick to, and make it work on Linux too. ;)
I know there's alternatives but nothing beats telling your coworkers "lets just use the built-in" when choosing a GUI stack.
My experience with Go is a while back, before it got generics. While I liked the overall idea of a language with little syntax and a focus on simplicity, I didn't really like Go in the end. It was a weird mix of too low-level and high level code. It exposed low level details that I don't need to think about in C#, and it is missing convenient abstractions like LINQ.
Personally, I like async/await in C# more than Go concurrency. I know this is heretical here, but I don't care about function coloring or any of the abstract advantages of other concurrency approaches. Async/Await in C# works for me (in non-legacy code) and is easy to understand.
I have also found c# async to be lighter than go's concurrency story, and dotnet11 is going to make that much more clear.
Now: I write software for fun, not for money.
I was on holiday, and stopped at a second hand bookshop to grab something hopefully a bit trashy to read. I picked up Nova Switch because the cover made it look like some weird sci-fi detective romp. But after about three pages it became apparent that I had happened upon something very special indeed, and when I made it to the end I did something that I'd never done before. I turned back to page one and read the entire thing again.
It is magnificent. Everyone should read it. It turned out to be the second book in a trilogy, and maybe that was part of its appeal. It didn't explain anything, it just launched into a beautiful and strange sci fi world, filled it with extraordinary characters, and moved languidly through the story.
So much science fiction is just dreadful garbage. Harrison is most definitely not.
Then, they essentially bankrolled their enemy by giving websites a way to monetize spammy content. By the late 2010s, they started losing the battle for many types of organic queries - although many HNers forget that the most common and "important" search queries are things like "Nike shoes", not "an implementation of a Bloom filter in Rust". And then, in the 2020s, they pioneered LLMs and made the technology public, and drove a final nail in that coffin by making it possible to produce SEO spam on any topic, including Bloom filters, practically for free.
Now, they see AI answers as their salvation, steamrolling a variety of non-monetizable use cases that never really mattered to them. And this will probably work only until AI-poisoning SEO is in full swing, at which point Google will find themselves with dissatisfied users and diminished ad revenue.
Which is one thing technologists often screw up: They misinterpret how people react to change; and they misinterpret how people really use their products.
Google Search used to be (and still is) a way to find places on the internet. As much as I like AI in searches, I agree; when I want to find the thing, I don't want AI getting in the way.
Media Molecule (LittleBigPlanet) spent over 8 years creating Dreams, which would never have happened without Sony's funding, but the tragedy was forcing what could have been a brilliant high-end authoring tool into a gamepad / lean-back interface. Even now, there are still innovative UX approaches in Dreams that could be adapted into high-end 2D, 3D, compositing and animation tools like Blender.
They have easy subscription options, even a monthly coffee adds up! Thanks!
Actually sorry not sure if they use triangle splatting, but gaussian splats at least.
My personal "AI Pro" Gemini has 3.5 Flash-Lite, 3.8 Flash, 3.1 Pro, 4 Argon.
Pretty inconsistent.
To add, it's actually surprisingly good. I have this unfinished paper draft that I keep feeding to various top AI's to see what it picks up on and what suggests to edit, and while Astra/Fable/Opus/Argon so far has been mediocre, Carbon's answer surprised me with some interesting points all the other AI's missed.
It's an anecdote, sample of one, but it got my attention.
They've been so slow to update their top models and roll them out that despite subscribing to the Gemini Pro plan, I use it so rarely that this is my last month. That and their product offering was so confusing I basically gave up.
Proofing and securing those is the way larger attack vector and work to come.
It's that, yes.
They didn't prove that the code has no bugs; they proved that none of the current bugs (known or unknown) are security-breaking bugs, per their formalized security model.
> First, our researchers collaborated with a mix of models to translate Signal’s security goals into formal mathematical specifications, with human experts reviewing every one. From there, agents powered by Gemini Carbon did the proofs, generating over six million lines of Lean, a computer language that mathematically proves the accuracy of code.
The generated Lean code here is just being checked against the "formal mathematical specifications of Signal's security goals", not checked for the low-level properties you might care about as a programmer, e.g. "will it ever leak memory" or "will it ever crash", etc. And in fact, the code is not even being checked for the reachability of non-security-violating protocol failure states, e.g. "could the protocol ever deadlock such that neither side can send messages any more".
It's still perfectly possible for the code to have all of these problems. It's now just known that, as of the commit they did this analysis against, none of those low-level problems, as far as they exist in the codebase, can cause a security-model violation.
Or, to put that another way: the system has been proven to "fail safe" in all its current failure modes, per their specific definition of safety (which is basically "keeping to the security guarantees Signal markets itself to users as having".)
I can believe that it's possible to get the LLM to write proofs, but I can't believe there were zero bugs in the system they verified. I wonder if they got the theorems wrong, or if they only proved narrow properties of the system.
I've had Claude formally verify crypto and data storage formats. It wrote theorems that looked fine and the formal verification passed but the resulting code contained many, many bugs that other implementations of the same things found with ad hoc testing.
The way the announcement is written, we should expect zero bugs in libsignal moving forward, and I guess time will tell whether that happens or not.
It's going to take way more than throwing a more advanced LLM at the problem to make any progress.
The extra velocity has meant we move considerably faster than the product team, so a lot of that extra dev capacity has been channeled into improving our tests, verifications and automated review processes - so hardly anything is slipping through the gaps.
This is the opposite of what I expected when we started working this way. One of the concerns we as a team raised was that it would lead to more bugs and lower quality but after 6 months the opposite appears to be true
I wonder how this agent is different than a classic coding agent, in the end it seems to work the same way, iterating over a codebase while writing Lean or Rocq code and running it.
Many puns and easter-eggs in both! Recommend giving it a quick glance, some are pretty fun. Some examples include 'fable' as a 'tail' (IE 'tale'), and 'universe' as a 'scale'.
Words can be present in both lists when the entity is applicable to both a tail and a scale, so a tailnet named `koi-koi.ts.net` is theoretically possible too.
Found out about it after I randomly lucked into rolling 'miku' as my 'tail' when first creating my account :)
Many other uses of tax dollars are handouts for a very small number of people that most of us see no benefit from (and may even have side-effects that make things worse for everybody else).
A beloved public institution, that improves society as a whole, is equally available to all, has enough money to do their mission well offer auxiliary services and to take advantage of PSA programs?
What a treasure.
The only permanent remedy is to outright ban those - or better: mandatory transparency into what politicians spend how much time with whom, where money streams come from and go out of organizations, etc, etc. All transparent in a public ledger. (and no: it is not 'anti-competitive' if it's the same regulatory landscape for everybody, but a way to safe-guard the government permanently from corruption). Of course if you there are ideas that would complete this or are truly overall better, feel free to throw them in the comments hat.
These people are a cancer.
This feels just so fundamentally like a step change to me. I was reverse-engineering Nest thermostats a year ago by hand, and now I feel like a month worth of my work could be done with 22 cents and an LLM. Exciting times for reverse-engineering and hardware/software interface
This, to me, shows a problem I have with AI: it is reinforcing inequalities. This is literally brute-forcing an OS by throwing money out the window.
It does not sound AT ALL like a clever way to build an OS. But they can do it because they have orders of magnitudes more money to spend on that than people who would actually know how to build an OS.
The thing is, brute-forcing software may be the best solution to make money. So if you are rich and can brute force software in a dumb way, chances are that you can make a business out of it. If you're poor but know what you are doing, you cannot compete.
Technically, it is impressive that they can get a program to do that, but they are wasting money and energy doing it.
As impressive as this sounds, whether it is tech teams or agent farms, the progress slows down considerably as the codebase grows. The real pain was seldom about shipping v1, but was more about how the v8 scales and how easy it is to get there, how robust the system remains between versions, and how constraints are held.
A huge part of the challenge is to pick/choose the design and implementation strategy that can only come from experience. That is one of the primary reasons why it's an absurd advice to me to let the llm do the goalseek and freestyle. This may be a no big deal on microservice or any managed runtime software that is an update away, but on an embedded or a high availability system where the field upgrade and hardware replacement costs could bankrupt a business.
if llm designed a circuit, will you bankroll it?
This is Bainbridge's 1982 paper Ironies of Automation restated.
https://en.wikipedia.org/wiki/Ironies_of_Automation
https://ckrybus.com/static/papers/Bainbridge_1983_Automatica...
My hypothesis - OpenAI and Anthropic are locked in battle. Whoever takes a step, the other one mirrors.
Based on this hypothesis, I think we can expect to see a new Astra model soon given the new Fable release.
And it could be that Anthropic released Fable 5.2 and kept 5.5 in reserve, in case the new Astra model performs better than 5.2
In particular, footnote eight:
> This is not a claim about the future. Perhaps almost all or perhaps even all of the software skills I have will become irrelevant in the near future. But we're not there yet. At the moment, programming skills seem higher leverage than ever before.
> There's something a bit funny about the current moment, where it seems like having the skills that models are replacing makes one much more effective than people who are trying to use models without having those skills. I like Max Bittker's framing that models are very good at counterfeiting and the level of expertise it takes to distinguish a counterfeit from the real thing has been increasing as models improve.
As well as this section fairly at the start:
> As far as I can tell, these problems are problems because people are implicitly carrying around a pre-agentic cost model. Ten years ago, the kind of person who implemented a new compiler because their compiler was slow was the kind of person who started programming a new 3d game engine when they wanted to make an indie game, i.e., the kind of person who never ships anything. Most productive programmers I know still have an aversion to this kind of thing, even though it is now actually reasonable to implement this kind of thing now.
Of course, at the moment everything’s still a bit very much difficult to tell, but “productivity”, at least in applicable areas of work, might really come down to who can let go of the most understanding the easiest? Incredibly thought-provoking stuff.
It has three different modes:
• You can simply describe a picture, and it'll automatically generate an image using procedural code.
https://kq-styles.specr.net/#/scene/custom/elk-valley
• You can ask it to paint over an existing image. That's what those examples show, like me creating EGA-style equivalents of various scenes from the original Myst.
https://kq-styles.specr.net/#/scene/myst/r01-c00
• The last way is to give it a rough pencil sketch. It's the mode I use the most since it's very valuable when building an adventure game to position objects in a room.
AI is actually still really bad at architecture. It produces very good localized code, like an auto complete. It produces so much of it, that it can be overwhelming, but it's not remotely better than the equivalent aka an infinite number of talented humans doing the work.
The scale of it (aka 'make me a video game') is daunting and I think misinforming our intuition as to what is going on.
LLMs are brute forcing outputs and attacking math problems is not a singularity, it's just more compute.
What we are witnessing is the ability for machines to use fuzzy concepts, instead of very strict code, which allows it to map into a much greater variety of things, aka code, pdfs, web - and crunch/compute the content.
It's just compute.
The "we don't read/write assembly anymore" argument doesn't fit though. Given an input x we know without a doubt we'll get output y. With an LLM, that isn't the case no matter how many guard rails, specs, markdown files, diagrams you give it.
What we're seeing instead is maybe akin to the green revolution. Before nitrogen fixation nitrogen was a resource constraining us humans. Lack of grit is what kept us from solving mathematical problems. However this lets us run into other resource constraints. A lot of technological progress does not solely hinge on thinking/processing. The speedup of creating new medicines with LLMs will probably be very small, as you still need to do all the testing. In software development you still need to first find out what to build in the first place and even mathematics where LLMs seem to shin brightest, it's becoming clearer that mathematics is not just solving mathematical problems.
Apart from that, a good coder is better than LLMs, shown by the amount of hand-holding needed. The only issue is that the amount of research an LLM can do and knowledge it is trained with far surpasses the amount of work humans would need to spend by hand, but we can also us LLMs for that.
I almost feel like saying, "we do like to build things that are impressive, but there's really only so much I can be impressed". These models have redlined my "impress-o-meter".
I'd like them to now cure cancer instead of giving me endless 3d worlds and crazy games, etc.
Interesting how low detail these were all up until xhigh.
https://tools.simonwillison.net/markdown-svg-renderer?url=ht...
max is in there too, but it ran out of reasoning tokens (128,000 limit) before it got to the actual SVG.
https://news.ycombinator.com/item?id=50422951
So only if you were "selected" for their experiments, your account was nerfed.
I’ve been shifting away from using more traditional GenAI models in favor of just giving LLMs access to a custom drawing language (LOGO-like) to generate imagery and its been going shockingly well - Opus 5.5 in particular.
I put together a simple LLM driven skill that creates old adventure game (EGA) environments from simple pencil sketches and it works really well. And being procedural, they're much easier to make surgical changes to without degradation.
Because on that specific benchmark (FrontierCode 2.0), Medium is Carbon's best effort level and xhigh is Fable 5.2's best effort level. You can see that on the graph your next point is about.
And it's just that one benchmark that's at medium. "All scores show each model’s highest-available effort unless otherwise noted".
> - How did they produce such a bizarre effort-intel curve for Carbon that goes against every available benchmark?
The graph is labeled as being about FrontierCode 2.0. Those appear to be Carbon's scores on the official FrontierCode 2.0 leaderboard: https://cognition.com/blog/frontier-code-2.0
> - Why did they leave Carbon-- positioned primarily as a cyber model-- off of the single cyber benchmark comparison?
> - Why do they switch to comparing against Argon (rather than carbon) for robodojo?
Because no numbers are available for Carbon for those benchmarks, and the model isn't publicly available so they couldn't run the benchmarks themselves?
And from your followup:
> where's Astra on that CVE chart
To run Astra on a vulnerability finding eval, you'd need to be part of OpenAI's verified cybersecurity program. Otherwise the results would be tainted by refusals. Do you expect that OpenAI would grant such access in this case?
I love seeing this verbalized. We are pretty much at the point that increased model capability will only have utility to folks that are willing to let go of even more understanding of whatever the model is producing.
The best models can/will produce a year's worth of output that would take several days to weeks to properly understand in the span of an afternoon.
For everyone that still wants to feel like they own the end result, just use the lesser models with less ambitious tasks.
Caveat: I think this mainly applies to coding. I would still like the best models possible to be working on e.g. new drug research to cure our various ailments.
Genuinely curious what that costs in terms of energy input and amortized computer cost vs. the previous version + a human's salary for 11 days.
- How did they produce such a bizarre effort-intel curve for Carbon that goes against every available benchmark?
- Why did they leave Carbon-- positioned primarily as a cyber model-- off of the single cyber benchmark comparison?
- Why do they switch to comparing against Argon (rather than carbon) for robodojo?
I've been using Opus 5.5 on xhigh for some decently aggressive software development on a ~125k-line project, and it has been fine for me. I'm not really willing to dive into a "rewrite the whole project, don't make any mistakes" level of refactoring. I'm not looking to try to prove Fermat's Last Theorem myself. Is there anything that I'm truly missing?
- Anthropic is finally well balanced across models. Meaning: Opus 5.5 is so good that I do not feel Fable FOMO as I did when it was Fable 5/5.1 vs Opus 4.8/5. And Sonnet 5.5 sometimes makes me forget about Opus.
- Fable is good, but you really have to push the ambition of your prompt in order to see proper separation from Opus 5.5. I personally have a hard time telling the difference between Fable 5.2 and Opus 5.5 for ordinary coding tasks.
- It’s way faster and more reliable: one of the worst things about fable 5.1 was how slow it was an, on top of that, how frequently it would suffer from max output token overflows. I got none of these issues with 5.2.
- Fable is exceptional at drawing. It’s ability to render an image with JavaScript actually has me questioning the value of dedicated image models. I prefer the images I’ve gotten fable/opus to make for me over those from gpt-image, it’s wild to think we’re at this point.
Eval context: 30 bug fixes and features evaluated by tests the agent didn't see. Ran on Opus 5.5 at medium effort in Claude Code.
20 of the tasks were last run on Nov 1, twice each. Nov 1 solved 36/40, and today solved 19/20. Thinking tokens per task were 0.95x Nov 1’s and output tokens 0.97x. The one significant change was speed: median time to first token decreased from 3.0s to 1.7s.
Caveats: one run per task today, medium effort only.
Opus 5.5 is a massive improvement.
Most people don't really have intuitions on what it means to push the frontier now (because AI has surpassed everything they are good at, for example, coding). We are truly inside the singularity (using Kurzweil's definition, slightly paraphrased - the time when technology progresses so fast no one can keep track with it).
Honestly incredible that they’ve managed to beat all benchmarks, while making it cheaper. Always happy to see improvements offered at lower costs.
Will be interesting to see what OpenAI does in response
Surprisingly the model is available immediately in my Claude Code: usually there's a delay, either for rollout or safety.
Was this before you lobotomized Opus 5.5 or after?
Having grown up in Ghana, the first game I tried was chaskele.
Was expecting a reference to BBC (Bela Boys Company) somewhere. Rather disappointed not to find one.
Never mind any issues about playability etc. Just the fact that AI has allowed someone to create a project that will make many people think a little bit more about Ghana make me happy.
Thanks for creating this and sharing it!!!
Mathias, if you are here and listening, I'm interested but $35 / month is the wrong price point for this. $420 for year.
I can buy DataGrip for $261 for three years and it has a perpetual fallback license that would let me pay for the first year and then stop and just not get updates.
IMO, to start, you want your product to be so good and so reasonably priced that it's easy for someone to decide to give it a try. You are coming in with a uber premium price and I think that's going to severely hinder your customer acquisition efforts.
Also, I almost never get a license to a tool like this where I'm not entitled to keep using it if I stop paying for it. There is just too much incentive on the part of the vendor to just keep driving the price up to the breaking point. With a fallback license like JetBrains has, if they get unreasonable about price, I can just exit.
Finally, in the age of LLMs, if this gets popular and the price point is wrong, someone is going to do a good job of producing a competitor and make it open source or much less expensive. Make the official version a reasonable enough price that there is no reason to not support you to make it better.
Just my $0.02 but hope it helps at some level.
Finally, I looked at the website and there was no way to contact you to tell you this personally. So...might want to fix that. Small org, easy contact should be part of the perk.
However, I wouldn't be willing to try it under proprietary-software terms. Last week, I was thinking about eliciting open-source software with the Ultorg look and feel from an LLM, thus replicating Bakke's HCI research and enabling other HCI researchers to build on it, which won't be possible under a proprietary license.
My notes on Ultorg may be of interest; you can
git clone http://canonical.org/~kragen/sw/pavnotes2.git
and read sieuferd.md (5200 words).The introduction:
> Eirik Bakke’s [SIEUFERD][0] (or, as he sometimes spells it, SIUEFERD) is a really interesting and promising generic database frontend, a “schema-independent database UI”. The elevator pitch is “a new kind of graphical query interface [with] the power of SQL but the ease-of-use of a spreadsheet”. (Unfortunately, he forgot to implement the DDL part of SQL, just doing the DML part.)
I can highly recommend Bakke's 5-minute talk on it from 10 years ago at https://vimeo.com/173726371.
I have previously read the printed version, so maybe now is a time for a very brief review. The book is technically fascinating and a worthwhile read, but it is also highly opinionated even by Lisp standards. Then again, it is refreshing in this time and age to see this deeply human and creative side of programming. I wholeheartedly recommend others to read it - and to carefully choose which of the advice to follow and which not :)
Thank you so much, Doug Hoyte, for making this book available to all of us!
A student came up to Von Neumann on campus and decided to prank the great scientist. "Two trains are on the same track 50 miles apart and they hurtle towards each other at 50 mph. As they start, a really fast fly on the nose of one train flies towards th eother at 60mph. When it reaches the other train it immediately flies back to the first, then back to the second, and so on until the fly is crushed in the crash. How far will the fly fly?
Von Neumann thought for a moment, and then replied, "30 miles!" The student grinned. "I thought you'd know this one! The trains take half an hour to hit each other, therefore the fly will fly 30 miles."
Von Neumann started. "I didn't think of that, I summed the series."
———
Source: "The Prisoner's Dilemma" by William Poundstone
https://archive.org/details/prisonersdilemma0000poun