Aaron Erickson

The amazing thing is a lot of this has been available from midjourney at this level for awhile, but the ChatGPT ability to distribute is making everyone rediscover image-gen again.

I heard Bluesky hates AI. Is that true?

This is kind of how malignant populism (aka Idiocracy) always tends to go. Cutting off your own dong to own the libs.

The good news is this bodes well for personal chicken coops once the price of an egg reaches the price of a bitcoin and a chicken becomes like a GPU

It’s dumb as shit on X and it’s dumb as shit here.

Wow, checking that out! Might see if it can do something interesting with scaled test time compute.

Idea: train an LLM on music theory and have it make music from first principles, using test time compute/inference to try things and test them out, maybe using RLHF at first to improve the model and iterate.

Anyone doing this? Would not only get around copyright, but might be more efficient.

Two narcissists married to each other can make it work, but I doubt Trump or Musk are willing to have wild make up sex with each other.

I remember when TSA did an experiment bringing out 3 lines, one for frequent travelers, one for families, and one for newbs.

We need a new one called "here is the line for instagram narcissists" staffed by the meanest, slowest person who got fired by the DMV for poor customer service.

Honestly, seeing guys like Greenwald and such go full horseshoe is probably the canary leading to some of these idiots going full MAGA just to accelerate thinking, delusionally, they will somehow emerge from "the revolution" victorious.

You're not getting denied jobs because you're white. You're getting denied jobs because nobody wants to work with assholes. Have you ever met a racist who wasn't also a massive asshole?

What good is a corner office if you can't walk by a bunch of small cubes on the way to it where all your worker bees are busy filing your TPS reports?

RTO has 2 primary backers:

- CRE who are taking massive losses on office buildings
- Execs who are big mad that the midlevel engineer that has better hair whose spouse is a professional decorator has a nicer office and ergo seems more important in zoom meetings

That’s a funny way to say bottle of pee

Anyone here using AI to reason through your company reorg? Having a lot of luck doing this for GPU jobs and clusters (i.e. which job should go in which cluster), seems applicable to that domain as well

There is going to be a lot of bullshit being sold in the next 2 years with the agent push, and a lot of careers are going to suffer when uninformed buyers spent lots of political capital going all in on broken platforms.

I beg... plead... please, if you are buying one of these, do some real research. Ask your real engineers what these things do, and take them seriously.

LLM demos are easy, prod is hard. Lots of startups now seem like they got a VC to see a demo, but did almost zero due diligence to see if it got part the hard technical parts that actually make or break success.

Its stupid stuff like access control, observability, getting different kinds of memory to work, managing each LLMs context window, etc.

It is things like understanding that having a set of agents that work with a common data source is 10x easier than very disparate sources where you need to translate terminology between domains. It is getting LLMs to understand that field X meant A after 2020, but meant B prior.

It is conversion of a subquestion to a contextually relevant query that understands your ubiquitous language (i.e. "what nodes are zombie nodes in our data center" - getting questions to use jargon effectivley is an actual hard problem).

The hard part of agents isn't the software glue to connect them. It's the prompts that allow a classifier agent to know what a specific agent can do and how to interpret the data.

Seeing an explosion of, for lack of a better term, "langchain wrappers" that are claiming to make AI safe for the enterprise.

Maybe some are good, but some of these are borderline fraudulent moneygrabs that are more of a sales/marketing operation than a technical one.

Convert them to nuclear and WV becomes the Silicon Hills for running training jobs on the us-west-virginia-42 cluster

I remember these guys from HS. They would break things for attention then, and only when nearly expelled, cry to the principal.

And why do PhDs get a special one? Time to bring back the Eng honorific if we’re gonna keep doing this

Funny how car drivers become "champions of banning able-ism" only when bike lanes or pedestrian improvements come up

"affordable housing"
"paycheck-to-paycheck"

Reading X is like talking with an LLM trained on the kind of people who you'd meet at the worst kind of HOA meetings.

Vain, mendacious, and not particularly insightful. You can literally predict the responses. Like a biological LLM trained on rhetorical poison.

Yes, I want something like Neuralink, connected to a system of LLMs, so I can just think "remind me of what X means" and have a nice voice in the back of my head give me the answer.

Ideally with an off switch for times I def don't need that voice in my head.