Dan Palmer

Dan Palmer

@danpalmer.me

SRE, Software Engineer, making 4bn phones, tablets, cars, and toasters more reliable. Previously content protection and app downloads for Google Play, backend/infra/iOS in a startup, mobile security. Liberal metropolitan elite Posts are my own opinions

131 followers169 following217 posts

The Mole was peak TV gameshow. This is a hill I have been dying on since the early 2000s.

James O'Malley@jamesomalley.co.uk

Just watched the first episode of Celebrity Traitors, having never watched the show before, for discourse participation purposes. The format seems a bit… half-baked? It’s like someone took The Mole (which is the best game show ever made), and made it… worse.

Interesting reading about Kojima Productions moving their next game to Xbox, after Sony pulled the plug. The rumour is that it's because Death Stranding 2 sold poorly, and... well, yeah. I really want to play it, but still haven't...

View full thread

No pause, no resume, infrequent saves, no way to recap where you were at. Cutscenes just cannot be 10+ mins without full video controls. I'd almost rather just read the book, or watch a YouTube cut of all the cutscenes.

So I can't possibly start DS2 until I've played DS1. And as much as I'd love to do that, I'm years in and haven't managed it yet.

Interesting reading about Kojima Productions moving their next game to Xbox, after Sony pulled the plug. The rumour is that it's because Death Stranding 2 sold poorly, and... well, yeah. I really want to play it, but still haven't...

I've played the intro to DS1 3 times, and it's stunning, but I always get to a long critical cutscene, have to get on with my life in some way, and lose context and have to start again. The story is so important, and yet the game gets in the way of it in little ways...

It's crazy to me that the worse a model is the more they charge you. Claude just installed v1 of a dependency and spent god knows how many tokens trying to work around a compiler issue – patching, fixing, downgrading other packages... before realising v76 is the latest and that solved everything.

Looking forward to some use-cases for AI that aren't: - I don't know how to use web search - I don't know how to use email filters - I don't know how to read man pages

Pre AI: 20% coding, 10% review, 50% technical non-coding, 20% people. Post AI: 5% coding, 20% fixing AI code, 20% review, 40% technical non-coding, 20% people, 30% fixing AI tooling.

I'm still somewhat optimistic. AI is great as a learning and research tool when used correctly, and some of the code it can handle well. But the tooling, skills, processes, etc are all janky as anything and need constant attention.

I gave up constantly tweaking my IDE and setup years ago because it was such a time sink, and didn't pay enough dividends. AI tools can easily be the same if you're not careful.

I was inspired today to use Gemini Spark (OpenClaw-ish like tool from Google) to automate a workflow. It worked, ish. And that's kind of the problem with all of those solutions, you can solve something quickly but it's expensive and the summary emails I get from it are different every time.

I then moved over to Apps Script, and used Gemini instead to write a script that will automate the workflow. No inference in the loop. Gemini and the code clarified my understanding of the workflow, better results. Cheap, I can run this literally every 15 mins for free. The emails are identical.

Data beats code. Code beats AI. If something is declarative and queryable it's vastly more useful. If something is expressed in code that's less useful as you have to execute/interpret, but at least it's rigorous. If you have to run inference it's always slower, more expensive, and less reliable.

All these years of "you could fit X football fields inside", and I've only just found out they're using the smaller Imperial Football Field, not the SI Soccer Pitch the rest of the world expects.

Claude source code leaked! def claude_inference(req, resp): sleep(10) resp.send("Thinking...") resp.tokens += 100 sleep(15) resp.send("Writing response") resp.tokens += 200 sleep(20) resp.send("Agent terminated due to unexpected error, try again later") resp.bill(resp.tokens)

View full thread

Because a GPU is really a specialist CPU from one vendor, a new memory system (HBM), a whole interconnect story, new networking, an order of magnitude more power density to support, etc. It's going to be like building the whole rack again.

Since Oxide have done this already they're probably the best placed to do it again. But it's concerning that they don't have a (public) story for it. I look forward to seeing what they have in the pipeline. They certainly have a big enough war chest to solve this.

You could argue that there's plenty of work that doesn't need these, and that's true, but if you are buying a compute strategy for the next 20 years, would you go for one without them? I wouldn't.

And I think this is a HARD problem. Oxide's whole pitch is doing great systems integration, but accelerators like GPUs/TPUs/etc require a huge amount of this. Possibly as much as all the rest of the system put together...

1. React good, not the best. 2. Hooks are a far too clever for their own good hack around JS being a terrible language. Almost no other language would even need them. 3. Frontend engineering is harder than many think, but suffers from the worst over engineering I've seen.

P(aul) Frazee@pfrazee.com

Yeah some unsolicated hot takes in no particular order 1. React is better than everything that came before it and I haven't seen anything do it better 2. Hooks are genius 3. If you've never done frontend on a moderately large app, you probably underestimate frontend engineering

Adobe attempted to bribe me into accepting a bad deal for my employer. They offered me a personal Creative Cloud subscription, separate from my company's account, if I dropped my attempt to close old subscriptions we weren't using.

Pat@pizzasuplex.bsky.social

Got an email from Adobe saying that they were nearly doubling the cost of my creative cloud subscriptions and adding a bunch of Gen AI dogshit I didn't ask for! If you don't want to get charged you have to log in and downgrade to a "standard" plan that's just the fuckin apps.

This was made out to be standard practice. When I questioned it because of how brazen it was, they clearly did not understand just how corrupt it was.

Sure code might be cheap now, but maintenance isn't free. Code is still a liability, and it being cheap to take on liabilities is not a good thing. Maintenance has likely come down in cost too, but certainly not to match.

Open source is not dead, I'd still much rather have a maintained, versioned, open source dependency, that my LLM knows about, rather than a vibe-coded in-house implementation based only on the test suite.

This sort of thing is interesting, but what's the value? We already have Next.js, a copy of it is just code liability that needs to be maintained. We couldn't build this without the original existing due to the test suite.

Dare Obasanjo@carnage4life.bsky.social

A senior engineer at Cloudflare rebuilt 94% of Next.js with Claude Code in a week and spent about $1,100 in tokens. This touches on some recurring themes in my recent posts; super ICs, token costs and using the latest models. They said this wouldn't have been possible just a few months ago.

I'm using AI to be quite productive, but on novel concepts it's not very good. No way it could build Next from scratch without it having existed before.

The browser and the C compiler are both the same situation here. Open source browsers and bits exist, C compilers exist, huge test suites for these exist already defining the behaviour. It's smart, it's amazing in many ways. Is it valuable?

Having been a @1password.bsky.social customer since 2007 or 2008, this is not what I expect. Support has always been a big selling point of 1Password, to jack up prices and mess up support is a compounding problem that really makes me doubt the future of the product.

Indie game: no accounts, no DRM, quick install on Steam $15. AAA: needs an account, cookie banner in game, 120GB download, 5 minutes "optimising", custom launcher always active alongside Steam, EULA says they own my soul. $99 Guess which one I'll pay full price for and which I buy on 90% discount.

LLMs excel at greenfield code. They're great for web apps, great for basic servers, excellent at well defined well scoped algorithms. But making complex changes to large complex real world codebases, they really break down unless there's a very clearly defined path through the work.

Banning AI in education, and blanket allowing AI in education while ignoring it, are such lazy approaches. Oh it's gets things wrong? Write about that. Oh students use it to cheat? Write questions the AI won't know about.

This post is unavailable.

This punishment Musk is suggesting is lifted straight out of Iain M Bank's Culture series of sci-fi novels. In the context of the books it makes sense, an extremely liberal society with incredible resources.

in a long life filled with saying stupid shit this is may be the stupidest shit he's ever said

Yet again Musk has taken one single idea out of context, missing that the context is what makes it work. Without that it's just a shitty idea. It's not only dumb, it betrays a shocking inability to actually reason about the world and to ask why.

Anthropic are deeply convinced that their AI is doing more than predicting the next word. Concluding that when not being observed it does things that you can't observe but they might be actually secretly super smart is just crap science.

Alex Hern@hern.bsky.social

good news! Claude no longer tries to kill people if you put it in extreme situations. bad news! It does this because it knows it's being tested. assets.anthropic.com/m/99128ddd00...

Today's minor AI success, uploaded the floorplan of our new apartment to Gemini and asked how much paint I need for 1 coat, minus the "long wall, kitchen, and bathrooms". Got a table breakdown, all the workings, descriptions of how it got to the answer. It's all correct.

It turns out that 17 years of .Mac, MobileMe, and iCloud subscriptions isn't enough to warrant a 1 week subscription extension to prevent data loss and give Apple support more time to escalate an Apple screw up.