Sebastian Broman

@thdxr And then there are the company behind one of the largest AI developer tools that does not even mention AI once

sometimes i get bored and go to a random tech company's website to see how they're pretending to be all about ai now

@jarredsumner Would be interesting to know the amounts of tokens used for such a project rewrite as well

there will be a blog post about this. on what this means for bun, benchmarks, memory usage, maintainability going forward, and also the literal process of doing this (it wasn’t just “claude, rewrite bun in rust. make no mistakes”) this is a 960,000 LOC rewrite, the code truly works, passing the test suite on Linux and soon other platforms. e2e I started working on this 6 days ago. this would’ve been a massive amount of work by hand.

99.8% of bun’s pre-existing test suite passes on Linux x64 glibc in the rust rewrite

@mariluukkainen Not an example from Finland, but it seems like loyalty programs doesn't have security in their top priority either. youtube.com/watch?v=O1zJ_0…

Finnish loyalty cards are one of the biggest jokes in the country. Here's an overview. S-Group and K-Group control over 80% of the grocery market. Nearly every Finnish household has an S-Etukortti, a K-Plussa card, or both. In most countries, shop staff ask if you found everything you needed. In Finland, the only thing you hear at checkout is "onko S-etukorttia?" The entire system is designed to make you feel like you're winning while you're being locked in. You get 1-5% back in bonus depending on how much you spend monthly. Sounds great until you realize that the duopoly sets the prices, so your "bonus" is a rebate on inflated pricing that exists because there's no real competition. In reality you're not saving money, you're just getting a fraction of the overcharge back and feeling somehow grateful about it. The genius is that the card is really not even a loyalty program. It's a big data collection machine. S-Group knows what you eat, when you shop, how much you spend on alcohol, which pharmacy you use, where you refuel. K-Group knows the same. They use this to optimize pricing, product placement, and marketing and give you back a few euros a month so you feel like you're getting a great deal. Lidl entered Finland without a loyalty card and just offered lower prices. They now have double-digit market share by doing literally the opposite of the Finnish model. Lidl's CEO publicly doesn't own an S-Etukortti. As a marketing strategy it works because deep down everyone knows the loyalty card game is rigged. The most Finnish thing about this is that people defend their card. "But I get 5% back!" Yeah, on groceries that cost 15-20% more than in Germany or Spain. The card is a psychological lock-in mechanism for a duopoly that has no incentive to compete on price because they've turned every household into a loyal data source.

@theo Would deno not be a more fair comparison with around 2400 unsafe blocks?

uv has 350k lines of Rust, and 73 "unsafe" calls. The Bun Rust port is already 681k lines of Rust, and has over 13,000 "unsafe" calls.

@DrizzleORM @saltyAom Maybe it does not cancel the query in Postgres, just don’t wait for the results?

@SebbeBroman @saltyAom I wonder how they do it, since pg protocol requires you to cancel query via a different connection

I asked about promise cancellation to stop running function in JavaScript once and they called me crazy because they thought it existed but it does not This is the only work around or roll your own fiber

@DrizzleORM @saltyAom #execute-and-cancelling-queries" target="_blank" rel="nofollow noopener">bun.com/docs/runtime/s…

@SebbeBroman @saltyAom Link plz

@DrizzleORM @saltyAom I believe bun sql has .cancel() on queries?

@saltyAom Yeah, no abort signals for pg drivers too unfortunately

@andrewingram Why either over lexical?

I like Tiptap, but every time I use it for anything serious I find myself thinking I should've just used prosemirror (especially with react-prosemirror being so compelling).

@pwign Is that some anti reverse engineering effort?

good news: got it working with capsules! bad news: opening chrome devtools will crash it

@jarredsumner Is this the bun bundler sourcemaps or does it apply to sourcemaps in general?

In the next version of Bun Source maps use up to 8x less memory

@cdngdev Just use qr-code-styling.com And separately some short url service if you need a smaller qrcode.

There is a special place in hell for QR code generator websites that silently replace your link with a redirect through their site, then break it after 10 scans and charge you to $35 reactivate it

@heliumbrowser It kind of makes me want to remove the show on hover and only using some hotkey or maybe just like a top corner to reduce the risk of "accidentally" showing it and displacing ui in the current page

@heliumbrowser Really like the zen mode. Unofficial tip, cmd+L selects the address bar insta showing it keeping it visible to read the url without hovering.

many new ways to make your browser feel "just right" dropped in helium today: - centered address bar - minimal address bar - new dynamic layout - improvements to vertical layout - and experimental zen/frameless mode (flagged, bit buggy, but cool!) helium.computer

@thdxr @ErdalToprak Is it usually public how long the cache time is? If a thread has gone “stale” it makes more sense to purge the history?

it's not even guaranteed not breaking the cache results in net cheaper costs never touching history means more context usage which we all know makes LLMs stupid ends up needing more compactions which end up needing more tool calls to rebuild up context we're trying to be more precise about this and share some data but we're fighting an uphill battle with people making unsubstantiated claims

Nobody talks about this but cache is literally holding us back If you could remove useless tool calls or dynamically change history tree there is so much perf to gain but it’s all about inference cost, for now

@qtnx_ @uwukko Then consider helium.computer/sponsor or would you only pay if other also had to pay?

@uwukko hope it works out, but just know i’d be the first customer for a 60 dollar paywall for helium more software should be comfortable with charging a one time price

maybe i should start selling a chromium browser with an ad blocker for $60 too

@maietta Very nice cache rate, but 884 GB in 107k requests, seems like you are serving very large payloads. Do you have big uncompressed assets?

Okay, this is terrible. This website is just a small blog. Put it behind Cloudflare and instantly in 24 hours burning through a lot of data and a huge number of requests. This has to be bot traffic.

@cszhu I guess this showed me my homepage does not give enough information about what it can do, I probably need to improve it a bit.

you can try it out here it's called "deathbyclawd": deathbyclawd.com let me know if there's any results that surprise you :-)

found a neat little website that gives your company a score based on how likely it is to survive the AI SaaSpocalypse. i think a lot of the companies we are familiar with today may not be here in 10-20 years... that being said... happy to announce that costco is immortal!! $1.50 hot dog forever!!

Saying thanks to a llm is basically tipping the provider

@DanielLockyer 40tok/s felt kind of fast before reasoning models

I'm hoping we can look back in 1 year and laugh at how we were all using models at 40tok/s Would love hundreds or thousands of tok/s

@dillon_mulroy My understanding is that last year much of the improvements came from scaling post training with more reinforcement learning. Optimizing attention would really help with longer contexts, since classic attention is quadratic.

it feels like the big break throughs with models has to be around optimizations with attention. is that right for those who are in the weeds on AI research?