pstoll

I’d argue we aren’t solving those inverse kinematics / kinetics via “number crunching” - but rather that our neuromuscular systems are analog. Which I don’t usually call that “number crunching” in the sense current computers … compute.

> I can practically see it playing out now: there is some huge incident of some kind, in some system or service with an MCP component somewhere, with some elaborate post-mortem revealing that some MCP server somewhere screwed up

Already happening.

https://www.infosecurity-magazine.com/news/atlassian-ai-agen...

> Falling back to CPU only

Patient: “Doctor, it hurts when I do this.”

Doctor: “don’t do that”

The other thing is that Microsoft does not own the copyright for any of the code they used. Facing their work on code they don’t own the copyright to is incredibly messy from an IP point of view.

It’s why con contributor licenses agreements exist in most open source popular projects.

Please tell me you are at least aware that tremor exists and that you rebuilt it on purpose?

https://www.tremor.rs/

> lol who is interested in this on hn?

I’m guessing lots of people who are on HN. I’ll go with - me for one. Guessing I’m not alone.

Remember - it’s a big world. Lots of POVs.

It was and always will be timsort-ly yours. iykyk.

Per second. Messages in the Kafka sense not the WhatsApp sense.

Meta (we) process many billions of messages per second. Finding enough randomness is a thing at scale.

Ah the "erlang = distributed" thing.

The primitives of sending messages across a "cluster" are built in the language, yep. And lightweight processes with a per-process garbage collector is magic on minimizing global gc pauses.

But all the hard work to make a distributed system work are not "batteries included" for Erlang. Back pressure on messages, etc. you end up havig to build it yourself.

We hit a limit at around 200 physical servers in a cluster back in 2015. Maybe it's gotten better since then. shrug.

As the author calls out, it was built for the 80s, when a million dollar switch shouldn't fail. It isn't built with https://en.wikipedia.org/wiki/STONITH in mind i.e. any node is trash, assume it will go away and maybe new ones will come back.

Rock on, Erlang dude, rock on.

When a sample showing a new SQL thing has eg

a) no id types eg for User, just a string b) uses i64 for “stars” C) any other of a million things…

I quickly conclude it’s likely a cute toy by someone who wanted to write a thing, not ever likely to be a battle-hardened effort brought forth by a high performance / scale database tortured soul. And I move on.

Yep - i thought / hoped / assumed it was about Optimizers aka Sat solvers, etc

A la https://github.com/google/or-tools

Also +1 for cromulent use of “cromulent”.

Having debugged my share of autoconf setups..I assumed it had to be a new M4 and not the ancient quirky GNU M4 thing because no one in their right mind would wish M4 (and related GNU autoconfig / autconf) on any other sentient beings.

It’d be like saying - “hey, I’m starting a new project and trying to pick between ed, sed, or awk. Whatcha think”? Def not.

It’s a question of risk.

Full disk encryption targets a different threat model - disk encryption protects against someone grabbing your computer.

Writing into an encrypted blob on disk adds a layer of protection against bad actors exfiltrating data by running code on the laptop.

Overall I really am amazed that this sort of thing is now possible and appreciate a privacy-aware / local compute and storage version of it!

Came here to mention Malloy. Which is from the team that built Looker, which Google acquired. The Looker CTO founder then started (joined?) Malloy. And … he recently 6mo ago moved from Google to Meta.

https://www.linkedin.com/posts/medriscoll_big-news-in-the-da...

Also for those playing along at home - a few other related tools for “doing more with queries”.

- AtScale - a semantic layer not dissimilar to LookML but with a good engine to optimize pre building the aggregates and routing queries among sql engines for perf.

- SDF - a team that left Meta to make a commercial offering for a sql parser and related tools. Say to help make dbt better.

(No affiliation other than having used / been involved with / know some of these people at work)

I also wonder what the perf overhead will be for programs that only need i32. I didn’t dig deep enough on the implementation, but these type of variadic runtime options often cause perf regressions because of all the “ifs” / lookups of type configs/parameters, etc. I just imagine some inner loops where mem.get adds a few instructions from the “if u32/64/“.

Unless it ends up being seamless / no cost via loading different u32/u64 implementations.

I mostly agree with the old c++ mantra of - no feature should have a runtime cost unless used.

If it ends up living up to the promise of the quality of the documentation (ie the README), I can’t wait to try it. Also screenshots of the display look cool.

> The work would be a multi-person-year effort, requiring concomitant funding.

> It is not clear to me how this work might begin.

To me that screams for - develop interest in a US public sector (aka government) funding agency and start getting grants to do this type of work.

I assume RISC-V can reasonably be pitched to government tech & policy folks as having many directly applicable benefits to their set of problems in a way they may want to invest in. To procure RISC- systems, they’d probably want a bunch of verification done. If the community isn’t investing, maybe they would.

Caveat I’m not current on the state of RISC-V adoption in public sector systems. But I’ve done funded systems development work at start ups with DARPA/ARPA, InQTel, etc. A motivated party should be able to make this happen.

Anyone closer to RISC-V know what the state of adoption in public sector?

I get where they are coming from though.

If the people building the product view building and using it as a chore, it’s going to show.

So I took “love” as their brand of Amazonian “customer obsession”.

There needs to be someone (often a founder, head of product, etc) who is the visionary for how the product can be awesome for customers. And hopefully that leader is able to infuse the entire product development experience with that energy.

The classic “big reveal” style of communication - hiding the important thing until half way through. Don’t bury the lede! It confuses and distracts the reader who is looking for how to frame and read everything else.

“SLC = simple, lovable, complete. Let me tell you why this is better than MVP.”

It ain’t hard, folks.

That cpu implementation is screaming for a cython implementation with all that bit twiddling. Would love to find a spare weekend.

Yep I know. It was easier to do in the same window with the program in the foreground vs opening another window to send the kill -HUP to this one. Double control-c still killed the process.

Random aside - I know formatting discussions border on the religious (and why something like gofmt is the only correct answer & yet I am also good with spaces for Python) but..

Did anyone else look at the code and ask themselves- what is the actual formatting standard being used?

Looked like a mix of “open brace on same line, 4 chars indent for code” then “open brace on new line, code at same zero indent”

Not a big deal obviously. Just something that tripped up my eyes scanning the code.

Cute, clever. There may be edge cases it doesn’t get right (parallel downloads/copies) but pretty useful when you’ve launched a job and didn’t think ahead of time to ask for progress.

I’ve used a signal handler (SIGHUP aka control-c) before as a “show progress” mechanism that I found very useful to monitor long running processes that were compute not io related (launched in screen fwiw to stay the active process).

It is the copy editors one ought to excoriate. If they still have any.

Neat. But I suggest moving the summary to the top.

There is a reason summaries / tldr are at the top. They help your reader understand / digest all the details that follow.

Have you ever operated a fleet of critical JVM instances and needed more memory? Don’t go over 32GB ram in an instance or the operating characteristics of you entire app change. Compressed memory pointers - oops. They are a blast to debug / operate!

https://stackoverflow.com/questions/25120546/trick-behind-jv...

Not paying too close attention, but this sounds a lot like the Gabor filter stacking from days of yore.

Ah looks like people are still looking at them. https://bihy.medium.com/improving-convolutional-neural-netwo...

You mean the “contributing” part of OSS?