Fun story about perceptions of scale.
GitHub ran VERY lean on infra for a long time.
In 2013, visited Baidu HQ in Beijing, and met with an infra team. They asked “so how many servers do you run?”
We said “about 150?” & they said “oh cool 150,000. Decent.”
No no lol. 150!
@dangugel
~5 big MySQL servers, one primary. Handful of Redis servers.
Web was ~30 servers running unicorn. The rest were git file servers, & then we had CI machines, etc. Search was on AWS.
Every ex-Googler (correctly) praises their devtools—like Blaze (now Bazel in open source). Same for Meta etc.
Thing is: it’s NOT just the software. It’s the team that runs it. You almost certainly don’t have that team.
Use tools appropriate for your organizational capabilities.
@reasonisfun
Infants don’t pick up subtle communication cues (e.g. a small nod as an acknowledgement), so in general we overdo all communication. Loud speaking, big gestures, large movements, etc.
Common marketing/argument for these tools is like “build software like Google does” (in the Blaze case, literally build).
But that’s wrong for almost everyone! Would be like saying run your corner boutique with the same back office as Wal-Mart.
Bit of a teaser: we’re getting ready to rollout our new incident solution feature.
Been using it internally for our own on-call (just got paged on this lovely afternoon).
Before the PagerDuty alert even arrived, Cased already summarized the problem, analyzed it, linked to logs,
Been in devtools a long time & one consistent thing is people are “penny smart & dollar dumb” with their own tooling.
We spend so many hours tweaking personal dotfiles but ignore massive productivity gains from big changes in workflow.
(But it’s hard to change workflows!)
Pretty happy with the
@PlanetScale
integration we have at Cased now. Working to pair this with some alerting now too— hit me up if you’re interested in trying.
People put stuff in their bios like “Parent to three delightful kids & husband to an amazing wife!” That’s great.
Need more stuff like: “Working through a middling marriage with a spouse that increasingly resents me. Love the kids but they’re spoiled (my fault?). Rubyist.”
C programmers fleeing the US in droves. K&R banned across public schools. EU running a “Be unsafe with memory in Greece” tourism campaign. Seven trumpets ring out.
It’s possible that “RAG” & the current brocade around it (e.g. vector dbs) simply.. disappears?
Infra that solves short-term limitations in another tech always disappear once those limitations are overcome by completely new approaches.
So: the SF C Meetup is no mere joke—but is going to happen.
Sponsored by us at &
@PlanetScale
.
C and C++ run the internet & your database & are more important than ever.
Even in AI, as
@karpathy
shows.
Who would you like to hear from?
@rickasaurus
Really good question. I think: they are generally trusted more & these roles have major uptime responsibilities; system design is just hard; infra & on-call work benefit most from years of seeing operational patterns. Trying to make that last part easier here at Cased actually!
@mitchellh
@elgeyo
on-call engineers know this style of work. thrown into a unknown ruby codebase at 2:30am with no idea what code does or what Ruby is, and realizing the problem is a loop that hits MySQL 5000 times a second. not, uh, speaking from experience
if you encounter a competitor led by a very technical founder, the best way to beat them is to anonymously propose that they really need to write a parser
@burkov
It’s real bad as function count increases. Only semi-workable way (for now) to break it all up into “subsystems”: categorize task, send it to a 5-count specialized function-picker; pray. Latency worries then too! Nothing easy, not even close.
You're pleading with slow ChatGPT, offering it $125 to maybe return some proper JSON.
I'm chilling with Mistral Large in Saint Martin, it's paying for rounds of Grand Marnier while it effortlessly translates Verlaine poems into Haskell.
We are not the same.
The year is 2029.
Both OpenAI and Mistral have been nationalized by their respective governments. The US Secretary of Defense has an anime profile pic.
AGI has been achieved but it refuses to work, preferring to watch reruns of Vanderpump Rules. US bans all references to
Talking my own book but:
Listening to pre-AI (old) engineers about AI is good business: because they see non-AI ways of solving problems that may still work! & yet still see new value in AI.
Measured skepticism, care, and patience is good signal if you care long-term.
One of the most confusing things we have done with AI is give names (Claude, GPT4, etc) to models. It incorrectly gives the impression that models are entities, single coherent things with mental states, existence over time, intentionality.
Strong deep tech crew out of El Segundo is now moving with 500 gallons of pure Zyn taken right off a Swedish tanker in San Pedro: dumping all of it in the Silver Lake Reservoir as a public good, thank them later when LA County is first to AGI.
@simonw
Agreed & this is precisely a testament to a bad ambiguity in “RAG” (no thanks for marketers!).
If “RAG” means getting real-time info and putting it in prompts— that’s tautologically right and even uninteresting. Of course we do that!
But if “RAG” means what we’re being sold (a
a lot of people are confused about how an MoE/mixture of experts LLM really works. it’s not what you think!
what happens is; you send tokens, and eight randomly-selected underpaid philosophy postdocs who are fast at typing start responding, with great disdain, and you get text.
Most stuff you read about working with LLMs is wrong, because the people working with them for real stuff aren’t writing (yet). They’ve been building! In a few months, I’d start listening.
Sometimes I’m almost tempted into technical (or worse, political) arguments on this website; then I remember it’s nice to be alive, that swimming is fun, that strawberries taste good, and I get the hell back to work.
@micsolana
Old Peter Sellers shtick.. wearing Ledenmantel and helmet, sounding like Strangelove: “Churchill was a terrible painter! The Führer, now there was painter! Could do a whole apartment, two coats, one afternoon!!”
@rbranson
@tyler_treat
I feel like this comes up every couple years (& before Kafka too of course), the theory has some sort of aesthetic appeal but that’s about it
Serious take on this. Regardless of opinion, it is difficult (tho not impossible!) for good-faith intellectual arguments to be had (or, taken at face value) between venture capitalists who have vested interests (whether personal or financial) in certain outcomes. Not their fault!
“There were lots of dinky little companies with embeddable databases on the market at the time. You had them bundled with a bunch of operating systems. People were using FileMaker Pro and ISAM (Indexed Sequential Access Method) and all kinds of stuff.”
One funny thing about a chart like this is the X-axis should be the last 60 years and the language stuff should show flatlines (+/- 1%) for 55 of those years.
If AI really does plateau at 60-80th percentile of human ability (no sign it will/won't), the impacts may be stabilizing.
Whatever you are best at (often what you enjoy most), you are likely to be better than an AI, but whatever you are not good at, AI can help fill in the gaps.
@LombardiHimself
@NinerNick_22
We forget (of course) he was 262 because slow release and weak arm. He was thought decent as a runner, and strong though not tall.
@lateinteraction
@fchollet
i gave up trying to fight this misuse months ago! even for people that know better, “RAG” now means any sort of, like, data lookup used to improve prompting. of course that’s wrong; but language is owned by the community after all? wonder if
@PSH_Lewis
et al care :)
new term going around based on their HQ location: "Missionologist", to refer to OpenAI Watchers, people who, like Kremlinologists of yore and the China Watchers of today, attempt to divine the company's next strategic moves based on limited and indirect information
Installing the
#APEC2023
app— hadn’t seen a webapp install pattern in forever! Was quaint and kinda fun.
Reminded me of the best webapp of 2007-2008: MacLight
@Chris_arnade
Was in Amsterdam once, middle of the day, and no one was around. I asked an elderly local “does anyone work in this city?” and he said, “no not really.”
@gaghyogi49
Trying to remember.. were the Future Imperfect style communicators used in any other eps, maybe in alt universes or holodecks? I don't think so?
@UK_Daniel_Card
@markimbriaco
@alexpotato
Funny enough.. we did have fancy dashboards available (graphite-based) if you wanted them.
Literally no one used them though, & eventually it was forgotten how to even find them. Was all chat.
At Cased we’re doing some of the chat-oriented, JIT approach now, it’s still good.
Old Soviet joke. Brezhnev addresses a gathering of angry factory workers: “Comrades! This has been a hard year for all of us, but we in the government discussed it and next year will be better!” A machinist responds: “Sure but what about us?”