By all means, keep using it, but it might be worth figuring out why other people don’t. Hint: it’s not because they’re more stupid than you or are looking to get promoted by big G.

(Personally, I like the ideas and binary encoding behind Capn Proto more than all the alternatives)

For high performance and critical stuff, SBE is much more suitable, but it doesn't have as good of a schema evolution story as protobuf.

Claude, ChatGPT, Gemini, and Grok are all more or less on par with each other, or a couple months behind at most. Chinese open models are also not far behind.

There's nothing inherent to these products to make them "sticky". If your tooling is designed for it, you can trivially switch models at any time. Mid-conversation, even. And it just works.

When you have basically equivalent products with no switching cost, you have perfect competition. They are all commodities. And that means: none of them can make a profit. It's a basic law of economics.

If they can't make a profit, no matter how revolutionary the tech is, their valuation is not justified, and they will be in big trouble when people figure this out.

So they need to make the product sticky somehow. So they:

1. Add a subscription payment model. Once you are paying a subscription fee, then the calculus on switching changes: if you only maintain one subscription, you have a strong reason to stick with it for everything.

2. Force you to use their client app, which only talks to their model, so you can't even try other models without changing your whole workflow, which most people won't bother to do.

These are bog standard tactics across the tech industry and beyond for limiting competitive pressure.

Everyone is mad about #2 but honestly I'm more mad about #1. The best thing for consumers would be if all these model providers strictly provided usage-based API pricing, which makes switching easy. But right now the subscription prices offer an enormous discount over API pricing, which just shows how much they are really desperate to create some sort of stickiness. The subscriptions don't even provide the "peace of mind" benefit that Spotify-like subscription models provide, where you don't have to worry about usage, because they still have enforced usage limits that people regularly hit. It's just purely a discount offered for locking yourself in.

But again I can't really be that mad because of course they are doing this, not doing it would be terrible business strategy.

If they're going to close the sub off to other tools, they need to make very strong improvements to the tool. And I don't really see that. It's "fine" but I actually think these tools are letting developers down.

They take over too much. They fail to give good insights into what's happening. They have poor stop/interrupt/correct dynamics. They don't properly incorporate a basic review cycle which is something we demand of junior developers and interns on our teams, but somehow not our AIs?

They're producing mountains of sometimes-good but often unreviewable code and it isn't the "AI"'s fault, it's the heuristics in the tools.

So I want to see innovation here. And I was hoping to see it from Anthropic. But I just saw the opposite.

I myself have been building a special-purpose vibe-coding environment and it's just astounding how easy it is to get great results by trying totally random ideas that are just trivial to implement.

Lots of companies are hoping to win here by creating the tool that everyone uses, but I think that's folly. The more likely outcome is that there are a million niche tools and everyone is using something different. That means nobody ends up with a giant valuation, and open source tools can compete easily. Bad for business, great for users.

I have no idea what JetBrain's financials are like, but I doubt they're raking in huge $$ despite having very good tools & unfortunately their attempts to keep abreast of the AI wave have been middling.

Basically, I need Claude Code with a proper review phase built in. I need it to slow-the-fuck-down and work with me more closely instead of shooting mountains of text at me and making me jam on the escape key over and over (and shout WTF I didn't ask for that!) at least twice a day.

IHMO these are not professional SWE tools right now. I use them on hobby projects but struggle to integrate them into professional day jobs where I have to be responsible in a code review for the output they produced.

And, again, it's not the LLM that's at fault. It's the steering wheel driving it missing a basic non-yeet process flow.

It sounds like you want Codex (for the second part)

It's irresponsible to your teammates to dump very large giant finished pieces of work on them for review. I try to impress that on my coworkers, and I don't appreciate getting code reviews like that for submission, and feel bad if I did the same.

Even worse if the code review contains blocks of code which the author doesn't even fully understand themselves because it came as one big block from and LLM.

I'll give you an example -- I have a longer term bigger task at work for a new service. I had discussions and initial designs I fed into Claude. "We" came to a concensus and ... it just built it. In one go mainly. It looks fine. That was Friday.

But now I have to go through that and say -- let's now turn this into something reviewable for my teammates. Which means basically learning everything this thing did, and trying to parcel it up into individual commits.

Which is something that the tool should have done for me, and involved me in.

Yes, you can prompt it to do that kind of thing. Plan is part of that, yes. But planning, implement, review in small chunks should be the default way of working, not something I have to force externally on it.

What I'd say is this: these tools right now are are programmer tools, but they're not engineer tools

I expect that from all my team mates, coworkers and reports. Submitting something for code review that they don't understand is unacceptable.

i immediately see that the most important thing to have understand a change is future LLMs more than people. we still need to understand whats going on, but if my LLM and my coworkers LLM are better aligned, chances are my coworker will have a better time working with the code that i publish than if i got them to understand it well but without their LLM understanding it.

with humans as the architects of LLM systems that build and maintain a code based system, i think the constraints are different, and that we dont ahve a great idea on what the actual requirements are yet.

it certainly mismatches with how we've been doing things in publishing small change requests that only do a part of a whole

Or to put it another way -- understandable piecemeal commits are a best practice for a fundamental human reason; moving away from them is risking lip-service reviews and throwing AI code right into production.

Which I imagine we'll get to (after there are much more robust auto-test/scan wrap-arounds), but that day isn't today.

In 2024, ~725 M$ total revenue, ~119 M$ net profit.

Well, no. It just means no single player can dominate the field in terms of profits. Anthropic is probably still losing money on subscribers, so other companies "reselling" their offering does them no good. Forcing you to use their TUI at least gives them control of how you interact with the models back. I'm guessing but since they've gone full send into the developer tooling space, their pitch to investors likely highlights the # of users on CC, not their subscriber numbers (which again, lose money). The move makes since in that respect.

Using openrouter myself I find the costs of APIs to be extremely low and affordable? I don't send the whole codebase to every question, I just ask about what I need, and everything is actually ridiculously cheap? $20 lasts about 3 months.

Meanwhile copy/pasting those shells in OpenRouter's Chat and asking the same question resulted in a single API request costing a tenth of a cent.

I could probably try tuning everything to keep costs down, but idk if it's worth the efforts.

I tried the same with OpenRouter and I used up 2.5 dollars in a day using Sonnet 4.5. Similar use on copilot has could maybe make me use 10% of my quota (and that's being generous for OpenRouter).

I think GitHub Copilot is way more affordable than OpenRouter.

I think it has nothing to do with the GPU and everything to do with the motherboard chipset.

https://www.capmetro.org/plan/schedmap?route=550

But much of the length is, in fact, single-track, making scheduling hard and meaning if a train is late or breaks down it disrupts the whole system.

And it's honestly pretty silly to see a train with the form factor of light rail but diesel-powered.

Voters did approve a proper light rail system in 2020 but it's gonna take a while to build and has already been scaled back twice, sigh...

I'm describing the red line which goes from downtown to Leander. (The one I linked to.) If you're describing some other line then sorry for the confusion, I didn't actually know there was another rail line. There are plans to build a green line to Elgin but AFAIK that's still under construction.

The red line is mostly single-track, but there are several specific segments of dual-track allowing trains to pass each other, which is why they're able to support multiple trains in both directions.

It's still a crappy schedule -- even during rush hour it's still no more than two trains an hour. Supposedly they intend to start running it every 15 minutes once they add some more dual-track segments.

Maybe it was worse when you lived here?

FYI Cloudflare didn't actually do that: https://x.com/eastdakota/status/1566160152684011520

(Disclosure: I work at Cloudflare but have no personal involvement with this.)

The "global protobuf db" had comments all over it saying it's not intended for production-critical tasks, and it had a lot of caveats and gotchas even aside from being built by Jeff's desktop, but it was so convenient that people naturally ended up using it anyway.

It was in 2008 I think (give or take a year, can't remember). I worked at Google at the time. Chunk Norris Facts was a popular Internet meme (which I think later faded when he came out as MAGA, but I digress...). A colleague (who wishes to remain anonymous) thought the idea of Jeff Dean Facts would be funny, and April 1st was coming up.

At the time, there was a team working on an experimental web app hosting platform code named Prometheus -- it was later released as App Engine. Using an early, internal build I put together a web site where people could submit "facts" about Jeff Dean, rate each other's facts on a five-star scale, and see the top-rated facts. Everything was anonymous. I had a few coworkers who are funnier than me populate some initial facts.

I found a few bugs in Prometheus in the process, which the team rapidly fixed to meet my "launch date" of April 1st. :)

On the day, which I think was a Sunday, early in the morning, I sent an email to the company-wide "misc" mailing list (or maybe it was eng-misc?) from a fake email address (a google group alias with private membership), and got the mailing list moderator to approve it.

It only took Jeff an hour or two to hack his way through the back-end servers (using various internal-facing status pages, Borg logs, etc.) to figure out my identity.

But everyone enjoyed it!

My only regret is that I targeted the site specifically at Jeff and not Sanjay Ghemawat. Back then, Jeff & Sanjay did everything together, and were responsible for inventing a huge number of core technologies at Google (I have no idea to what extent they still work together today). The site was a joke, but I think it had the side effect of elevating Jeff above Sanjay, which is not what I intended. Really the only reason I targeted Jeff is because he's a bit easier to make fun of personality-wise, and because "Jeff Dean Facts" sort of rolls off the tongue easier that "Sanjay Ghemawat Facts" -- but in retrospect this feels a little racist. :(

My personal favorite joke is: Jeff Dean puts his pants on one leg at a time, but if he had more than two legs, you'd see his approach is actually O(log n).

But I'm fully aware you wouldn't want a "Sanjay Facts", and that's not the point. ;)

Legend. Popping up here after his last comment was 13 years ago.

This web app was also how I learned the pain of maintaining a live web service with a lot of ever-changing dependencies. How I sighed when the AppEngine version changed and I had to fix things again...

I handed it off again before I left Google but I have no memory of who that was to unfortunately :(.

Thanks so much for falling for my trick and taking it over, I was getting pretty sick of dealing with the same issues you describe. :)

One of the reasons Cloudflare Workers has a policy that we absolutely never break deployed code, ever. (Well... at least not intentionally...)

But in a more literal sense, the chance of your joke landing was likely higher due to the things that you stated and due to your audience and their biases.

I don’t see your joke as being in any way harmful towards Sanjay aside from potential knock on effects of Jeff Dean being more popular. But if you try to calculate every second and third order consequence of everything that you do, let alone any moments of humor you might have.. Well, you might as well lock yourself in a cell now.

I mean… yeah. When two people are peers and comparably well regarded, and one is elevated above the other and enjoys increased popularity, familiarity, and respect, and the elevation is because that person's name comes from a culture that is more aligned with the dominant culture and easier for them to engage with… that is a pretty textbook example of systemic racism.

I'm not at all saying this to demonize Kenton. We can make mistakes and reflect on them later, and that's laudable. But it is important to recognize these systems for what they are, so that we can notice them when they happen all around us every day.

We all have cultural biases and familiarities, is that wrong? By this definition, we're all racist. Maybe that's true but it kind of ceases to be a useful distinction at that point. I wholeheartedly agree with your last sentence, but I don't know if throwing around the r-word is helpful.

Does it? I would argue that recognizing that we all swim in a soup of cultural biases and familiarities that advantage some people and disadvantage others is a noteworthy insight, an insight with practical implications. After all, we aren't volitionless molecules bouncing off walls. What if we made an effort to observe these biases more closely, to study there effects, and to better understand the way they effect our own behaviour? Then, what if we made an effort to counteract these biases, first in our own behaviour and then in our communities?

Are we not?

The free will debate aside, I think what you said makes a lot of sense, and comes across as empathetic, and you didn't need to use that word. I just think it's too loaded, aggressive, and broad to be very useful as a shorthand for the more complex thought you expressed.

I looked at it from the perspective of syllable count. Jeff Dean is easier to say by that measure. If Jeff were instead named Alexander Chesterton, would he still be the obvious choice to head the facts? My takeaway from this is that a single-syllable name is perhaps a great boon.

I'm not calling myself Hitler here, I think it was a mild offense. But in retrospect the site could have been about both of them, with competing facts, and that could have been really cool. Oh well.

It’s not “racism.” There’s plenty of Indians with names that are easy for English speakers. Conversely, the same situations would’ve presented itself if the other person was any sort of white Eastern European.

In fact, calling this “racist” is itself racist. I have close friends with family names from Poland or Croatia where we don’t even try to pronounce their names correctly. Nobody feels bad about that. But for some reason if it’s a “brown person” we’re suddenly super sensitive about it. That is differential treatment based on race.

People get awkward about how to pronounce my name because I’m brown. But it’s hard to pronounce because it’s misspelled Germanic! They wouldn’t act that way if I was a white guy with the same name.

As the world's foremost expert about what happened in my head, do I get to, like, pick a winner here?

If so I pick tczMUFlmoNk, I think their description is accurate. (I think you might want to re-read it as it feels like you are responding to something else.)

If I don't get to pick, this is quite weird! "People on Hacker News tell me I'm wrong about my own thoughts." was not on my -- actually wait, that doesn't sound unexpected at all now that I write it out! OK, carry on.

1. The original choice: Kenton picked "Jeff Dean" because the name was more familiar/rhythmic in English. This wasn't about skin color, it was about name patterns. You're right that a Polish surname could have the same issue, and in that, you're demonstrating complete understanding of the issue at hand.

2. The reflection afterward: Recognizing that name-familiarity advantages systematically correlate with certain cultural backgrounds more than others isn't "differential treatment based on skin color", it is observing a statistical pattern in outcomes.

And here's the key point: given Kenton's explanation, they are indicating they would reflect the same way if Sanjay had been Polish with an unfamiliar surname. You're arguing with Kenton about what Kenton thinks and could think... while Kenton is right here. At some point you have to engage with what he's actually saying rather than insisting you understand his mind better than he does.

(Which again to be clear I'm am not riven with guilt here, I just think maybe given another chance I would have made it about both of them.)

Obviously what’s in your head specifically is idiosyncratic to you. But the feeling you’re having certainly happens more generally, and is based on general social understandings. That’s what I’m commenting on.

If I say, “this feels a little rude,” isn’t it fair for people to chime in as to whether it’s actually rude by reference to general social standards?

For what it's worth, I personally regard Sanjay in just as much awe as Jeff and understand that the meme is just an Internet meme and nothing more.

You need to take a breath, read what people write, and stop trying to win the argument.

Mr Rayiner is a lawyer by profession ;) https://news.ycombinator.com/item?id=11340543

https://www.newyorker.com/magazine/2018/12/10/the-friendship...

EDIT: And HN at the time:

https://news.ycombinator.com/item?id=18588697

Jeff always had an idea how to make something a bit faster using a clever trick, but Sanjay would respond by coming up with how to express the abstraction in a way that other mortals could comprehend, or just telling Jeff it wasn't a good idea because it would make things more difficult to maintain.

Jeff was also prone to dad jokes, Sanjay's humor was far more subtle. Both were awesome to talk to and one of my proudest moments was when Jeff read a document proposal I wrote ("Google should get involved in Genomics Research and Drug Discovery") and took it seriously.

Verily was its own thing that was unrelated to my push in Research. I think Larry Page knew Andy Conrad and told him he could do what he wanted (which led to Verily focusing on medical devices, which is a terrible industry to be in). They've pivoted a few times without much real success. My hope is that Alphabet sheds Verily (they've been trying) or just admit it's a failure and shut it down. It was just never run with the right philosophy.

Calico... that came out of Larry and Art Levinson- I guess Larry thought Art knew the secret to living forever and by giving him billions Art would come up with the solution to immortality and Larry would have first access to it. But they were ultra-secretive and tried to have the best of both worlds- full access to Google3 and borg, but without Googlers having any access to calico. That, combined with a number of other things, have led Calico to just be a quiet and not very interesting research group. I expect it to disband at some point.

Isomorphic is more recent than any of the stuff I was involved in, and is DeepMind (specifically Demis's) attempt to commercialize their work with AlphaFold. However, everybody in the field knows the strategy of 1. solve protein structure prediction 2. ??? 3. design profitable drugs and get them approved... is not a great strategy because protein structure determine has not ever been the rate limiting step to identifying targets and developing leads. I agree I don't really see a future for it but Demis has at least 10-20 years of runway before he has to take off or bail.

All of my suggestions were just for Google to do research with the community and publish it (especially the model code and weights, but also pipelines to prep data for learning) and turn a few of the ideas into products in Google Cloud(that's how Google Genomics was born... I was talking to Jeff, and he said "if we compress the genome enough, we can store it all on Flash, which would make search fast but cheap, and we'd have a useful product for genomics analysis companies"). IMHO Jeff's team substantially achieved their goals before the DeepMind stuff- DeepVariant was well-respected, but almost every person who worked on it and related systems got burned out and moved on.

What is success, anyway, in biotech? Is it making a drug that makes a lot of money? What if you do that, but it costs so much that people go bankrupt taking it? Or is the goal to make substantial improvements to the technology, potentially discovering key biological details that truly improve people's lives? Many would say that becoming a successful real estate ownership company is the real destination of any successful pharma/biotech.

In my opinion forays into biology by moonshot hopefuls fail for one of two reasons: either they completely ignore all the current wisdom from academia and industry, or they recruit the very academia people who are culturally responsible for the science rot we have at this time. Calico (and CZI, and im starting to fear, Arc) fell prey to the latter. Once you recruit one tenured professor IMO youre done. The level of tenure track trauma and academic rot they bring in can burn even a trillion dollars into dead-end initiatives.

IMO (after decades of daydreaming about this scenario), the only plausible way to recreate a Bell labs for Biology is to start something behind a single radical person, and recruit the smartest undergrads into that place directly. Ensure that they never become experts at just one thing so they have no allegiance to a method or field. And then let that hoarde loose on a single problem and see what comes out. For better or worse neuralink seems to be doing that right. Just wish they didnt abuse the monkeys that much!

To me success in biotechnology is if I measurably help make a drug that makes a person smile and breathe easy that would otherwise not have. Surprisingly easy and hard at the same time.

It’s not racist. It’s just to do with name length.

:)

They just got their own (unofficial) Lego set: https://x.com/JeffDean/status/2006581022666928415

Later version of the site was generalized so that people can submit facts for any user. I think Jeff Dean still has all the funniest fact though.

my nonexpert impression is jeff keeps much more of a public profile. hence the natural celebrity goes to him. was this not true way back in the day?

With that said, I suppose it's hard for me to say what the public perception of the two was in 2008 as I only knew of either of them from working there.

You still have to create a Youtube channel for Jeff Dean/Sanjay Ghemawat (slayers of code!) like "Entertaining AI" did for "Chuck Norris" - https://www.youtube.com/@Entertaining_AI

You could add a meme generator that's like the Django docs tutorial with the internet web-poll

Because it’s easy to pick on myself. :)"

Damn!

The reason it rolls off the tongue easier is because of the familiarity with names of that form. It is making a choice to favour a personal cultural similarity. It clearly wasn't done with malice, but being reflective about it and noticing that it happened is a good thing.

When people talk about privilege, this is a large part of what people mean. There's no intended bias, it is just an honest choice, but all of our choices are based upon our opinions that will inherently have biases of some sort. One factor of privilege is when those choices disproportionately fall your way because decisions end up being made by people who you share a cultural upbringing with. Their intuitive decisions value what you value.

Sometimes the only way you can deal with that is by acknowledging your intuitions contain that implicit bias and dispassionately try and balance them as best you can.

I’d be interested to hear from someone from a different culture to verify whether this is true or not.

It's not an absolute claim being made here. It's just a consideration that what we intuitively feel may not be an expression of a universally held value.

workerd does not include any sandboxing layers other than V8 itself. If someone has a V8 zero-day exploit, they can break out of the sandbox.

But putting aside zero-day exploits for a moment, workerd is designed to be a sandbox. That is, applications by default have access to nothing except what you give them. There is only one default-on type of access: public internet access (covering public IPs only). You can disable this by overriding `globalOutbound` in the config (with which you can either intercept internet requests, or just block them).

This is pretty different from e.g. Node, which starts from the assumption that apps should have permission to run arbitrary native code, limited only by the permissions of the user account under which Node is running.

Some other runtimes advertise various forms of permissions, but workerd is the only one I know of where this is the core intended use case, and where all permissions (other than optionally public internet access, as mentioned) must be granted via capability-based security.

Unfortunately, JavaScript engines are complicated, which means they tend to have bugs, and these bugs are often exploitable to escape the sandbox. This is not just true of V8, it's true of all of them; any that claims otherwise is naive. Cloudflare in production has a multi-layer security model to mitigate this, but our model involves a lot of, shall we say, active management which can't easily be packaged up into an open source product.

With all that said, not all threat models require you to worry about such zero-day exploits, and you need to think about risk/benefit tradeoffs. We obviously have to worry about zero-days at Cloudflare since anyone can just upload code to us and run it. But if you're not literally accepting code directly from anonymous internet users then the risk may be a lot lower, and the overall security benefit of fine-grained sandboxing may be worth the increased exposure to zero-days.

The problem I have is that I'm just one person and I don't want to be on call 24/7 ready to react to sandbox escapes, so I'm hoping I can find a solution that someone else built where they are willing to say "this is safe: you can feed in a string of untrusted JavaScript and we are confident it won't break out again".

I think I might be able to get there via WebAssembly (e.g. with QuickJS or MicroQuickJS compiled to WASM) because the whole point of WebAssembly is to solve this one problem.

> But if you're not literally accepting code directly from anonymous internet users then the risk may be a lot lower

That's the problem: this is exactly what I want to be able to do!

I want to build extension systems for my own apps such that users can run their own code or paste in code written by other people and have it execute safely. Similar to Shopify Functions: https://shopify.dev/docs/apps/build/functions

I think the value unlocked by this kind of extension mechanism is ready to skyrocket, because users can use LLMs to help write that code for them.

For Wasm to be a secure sandbox, you have to assume a bug-free compiler/interpreter, which, alas, none of them really are. It's a somewhat easier problem than building a bug-free JavaScript runtime, but not by as much as you might expect, sadly.

> I want to build extension systems for my own apps such that users can run their own code or paste in code written by other people and have it execute safely. Similar to Shopify Functions: https://shopify.dev/docs/apps/build/functions

Ah, this is exactly the Workers for Platforms use case: https://developers.cloudflare.com/cloudflare-for-platforms/w...

And indeed, Shopify uses it: https://shopify.engineering/how-we-built-oxygen

(There's also the upcoming Dynamic Worker Loader API: https://developers.cloudflare.com/workers/runtime-apis/bindi...)

But it sounds like you really do want to self-host? I don't blame you, but that does make it tough. I'm not sure there's any such thing as a secure sandbox that doesn't require some level of monitoring and daily maintenance, sadly. (But admittedly I may be biased.)

I've been picking at this problem for a few years now!

On the one hand I get why it's so hard. But it really feels like it should be possible to solve this in 2026 - executing arbitrary code in a way that constrains its memory and CPU time usage is a problem our industry solves in browsers and hosting platforms and databases and all sorts of other places, and has done for decades.

The whole LLM-assisted end-user programming thing makes solving this with the right developer affordances so valuable!

If Simon's users choose to self-host the open source version of his service, they are probably using it to run their own code, and so the sandbox security matters less, and workerd may be fine. The sandbox only matters when Simon himself offers his software as a service, which he could do using Workers for Platforms.

(But this is a self-serving argument coming from me.)

Benchmark numbers for request isolated JS hello world / React page rendering:

    JCO/wasmtime: 314µs / 13ms
    Bun process forking: 1.7ms / 8.2ms
    v8 isolate from snapshot: 0.7ms / 22ms
    TinyKVM: 52µs / 708µs
    Native with reuse 14µs / 640µs

(I suspect even if we are able to get TinyKVM into a state you'd feel comfortable with in the future it would still be an awkward fit for Datasette since nested virtualisation is not exposed on AWS EC2.)

Even if you go with something backed by a full time team there is still going to be a chance you have to deal with a security issue in a hurry, maybe in the run up to Christmas. That is just going to come with the territory and if you don’t want to deal with that then you probably need to think about whether you really need a sandbox that can execute untrusted code.

It hasn't been updated in some time, but it should still be working, and can probably be brought up to date with some small effort: https://github.com/facebook/hermes/tree/static_h/API/hermes_...

EDIT: Reading some of your other comments, I should point out that this is more like a component of a possible solution. It does not attempt to prevent resource exhaustion or crashes due to corrupted internal state.

https://developers.cloudflare.com/sandbox/