Tyler Ham

For yesterday & today's SWE work, I agree. I can envision a future where we're asking the models a lot broader, longer spanning questions where they can work a lot longer, "wider" and have broader responsibility to develop larger systems. Then we get humans and cheaper models to go thru & review.

Ed@ed3d.net

This is where benchmarks are going to start making progressively less sense in general. Like, on a pure happy-happy-joy-joy code-grind basis, 95% of developers might do something that would represent the top half of DeepSWE...what, once every five years? We have saturated HumanDevBench thoroughly

16 Sept

Without data, marketing is even more faff and fluff than usual. Especially without any reliable standardized 3rd party way of comparing two products, even if they are two products from the same company.

Juanmi@juanmirod.bsky.social

If the benchmarks that we have are saturated... What are they using to train the models? How do they know if they improve? They should be much more transparent about the training data and methods and we would know better also what they can really do.

10 Sept

This is the way.

oliver@eikopf.com

maybe this is a universal truth i’m only discovering now, but working on internal tools is way more fun than doing literally anything else

This regulation already exists, which is why if you look closely, the only claims are unfalsifiable. Full of testimonials which may be true individually. "Feels", "helps with" are difficult to falsify, even 0.01% can be true. The marketing has found the holes and aligned with the existing science.

ghost malone ergo propter bones@proptermalone.bsky.social

does your supplement do what you say it does? cool, prove it. if you can’t prove it and you keep saying shit you can’t prove, fines and jail.

You still can't tell if something will work for you without trying to see if it matches your individual conditions and biology: i.e. you could be one of the outliers of the averages.

Anyway, this is the self rationalization I go thru to justify trying all this stuff 🙃. This is my personal bias. I find it a fun personal hobby, and try to spot marketing to steer others away from fakery.

ghost malone ergo propter bones@proptermalone.bsky.social

does your supplement do what you say it does? cool, prove it. if you can’t prove it and you keep saying shit you can’t prove, fines and jail.

The next level is celebrity endorsements. Also testimonials. It'd be difficult to stamp out.

Plus, scientific studies are averages drawn from a pool of "regular people" or people with specific conditions. You may be looking at data from one group but you are in the other group.

I should start posting on my own timeline. Here's a start: I enjoyed seeing some of you at the ATmosphereConf in Vancouver back in March.

Tyler Ham reposted

Doll

@dollspace.gay

29 Aug

If rising productivity doesn’t buy ordinary people more time and resources for self-directed creation, curiosity, mastery, and play, then what the fuck was the productivity for? This graph is considerably more encouraging than “luxury consumption rose.”

Tom@moll.dev

wild chart

1 Apr

Needs to be seen on the site, has a wonderful drop-down menu.

Randall Munroe@xkcd.com

CREATION Happy April Fool's Day! xkcd.com/3227/

13 Mar

In addition to AI writing tests, it's also easier than ever getting to or near 100% code coverage.

hikikomorphism@hikikomorphism.bsky.social

I have an idea for this: adverserial mutation, just spawn a bunch of cheap Haiku agents and have them deliberately introduce bugs at varying levels of subtlety, every bug that the preexisting test suites don't catch signals a gap in test suite coverage

8 Mar

Question: Some SWE jobs/companies are really protective of the code they produce and protect it with copyright including if an employee were to leave a job or work for a competitor. What happens when everyone expects an AI to write code and the USPTO says the code is no longer copyrightable?

insideevs.cominsideevs.com

26 Feb

I can think of only one legitimate reason to do this, which aligns with the statement that not many people used it: If the assembly of the Mach-E now excludes the plastic liner and assembly of the frunk there's a little bit (not $500) of savings.

e.w. niedermeyer@niedermeyer.online

holy shit, the car business has absolutely lost the plot insideevs.com/news/788178/...

youtu.beYou wouldn't STEAL a CAR? (Hilarious anti piracy commercial)YouTube video by PopMov

24 Feb

We need to bring back the term freebooting for copyright infringement. Model training and distillation is freebooting and TOS violation but not theft unless you agree with the "you wouldn't steal a car" anti-piracy ads from years ago. youtu.be/PLzy-IJg4_4

28 Jan

I'm amused that the people talking about this are using this app, the phones and OSes which all have code written with LLM assistance.

Ryan "UnionToTuple" Cavanaugh@searyanc.dev

Getting tired of hearing "There are literally no use cases for LLMs" from people who think the only way to use an LLM is to type into chatgpt dot com, press enter, and hope for the best

21 Jan

One of the things that I learned as a SWE consultant taught me is that the code you write will always eventually be owned by someone else. Your time there will come to an end. Therefore the code has to not only be readable & understandable but the sooner the entire team owns all the code, the better

Steve G@stevewithag.bsky.social

Actual Engineering Manager here. Everyone talks about Vibe Coding, but I don't hear a lot about Vibe Debugging. My nightmare is that something burps out a mess of code and *no one is responsible for it*. If one of my people writes it... well they get to fix it. If no one writes it.......

21 Jan

Globally distributed teams have similar considerations: when my coworkers in India write code, I will be working with it and adding to it hours later while they sleep. And vice versa.