I relate to the points about obscure platform limits, and leaky abstractions, but when I look at the exe.dev platform, it might be the most obscure PaaS I've seen, and has it's own strange abstractions.

The shell command to start a new vm, has a --prompt flag to get an LLM to configure the VM for you.

VM's have no public ipv4 IP, and the ipv6 IP doesn't seem to allow incoming connections.

The only supported inbound connections are via their HTTP proxy.

There is no private networking.

At first I interpreted the complaint about cloud providers not offering nested-virtualization, as something he intends to address by offering it as a feature, but no, instead he means that exe.dev's VM abstraction eschews the need for it.

Yeah I had a giggle about that also. He argues: “cloud abstractions are the wrong shape”, then what they actually ship is: a different abstraction, with even more hidden constraints.

I'm very curious how they deal with subscription levels/noisy neighbors.

How do you collect these stats?

Is it by characters human typed vs AI generated, or by commit or something?

> How do you collect these stats?

Cursor dashboard. I know they're incentivized to over-estimate but feels directionally accurate when I look at recent PRs.

All of their benchmarks are against 16 bit models right?

Why aren't they comparing to 2/3/4 bit quants?

looked at quant versions of these models and they all outperform it so I guess it just doesn't look as good.

It's also just wrong, their SQLite benchmark is only using a single thread for the SQLite connection. It's much faster with multiple connections.

I installed two models, but it just always reports:

    Available models (2):
    CohereLabs/cohere-transcribe-03-2026 (4.6 GB)
    flux_2_klein_9b_q8p.ckpt (20.2 GB)
    ...
    Advertising 0 model(s) (only loaded models)

Also the benchmark just doesn't work.

Interesting idea, but needs some work.

The sqlite benchmark is not very optimised.

Just adding:

    db.SetMaxOpenConns(runtime.NumCPU())
    db.SetMaxIdleConns(runtime.NumCPU())

Made the performance on my machine go from 27,700.99 r/s to 89,687.36 r/s.

I also tried making the get user a prepared statement, and storing the timestamp as an unix timestamp integer, but that didn't make much difference for me.

He bailed out his Twitter investors by having xAi purchase Twitter, then he bailed out his xAi investors by having SpaceX purchase xAi, and how he is trying to bail out SpaceX by having the index funds be forced to purchase SpaceX.

Twitter valuation aside, there must be some intangible benefits to the purchase. Lots of influential people use Twitter, he has amplified his reach, and the purchase seems to have moved the overton window in favor of his agenda.

> the index funds be forced to purchase SpaceX

SpaceX is a fantastic company full of the smartest engineers that has brought fast internet to the most remote regions on Earth. Even if it wasn't for index funds, investors would buy the shit out of it.

xAI has built an almost-frontier model. It's not Anthropic or OpenAI but it's also not valued like Anthropic or OpenAI either.

Twitter is doing OK, despite predictions. It found a niche with obnoxious right-wing assholes instead of obnoxious left-wing assholes. Twitter does seem to be less excited about getting people fired for wrongthink than before which is a societal improvement. It did succeed in buying Musk influence at the highest levels, but he threw it away. Mixed success.

I know Musk gets a lot of hatred on here for his behavior and beliefs, including from myself, but that doesn't imply that his businesses are all scams. Hugo Boss was a Nazi too and his business is doing fine.

Elon has made multiple comments in the past about how much keeping SpaceX private, makes it easier to run, and achieve it's goals of colonising Mars.

He has given up on that, he has abandoned the Mars goal, and he is sacrificing the company to the whims of the public markets, just so he can have a payday.

xAi raised $45 billion and Grok loses benchmarks to Chinese(GPU sanctioned) open weight models.

I'm overall positive about the Twitter acquisition, I mean it was terrible financially, and Elon has made a lot of unforced errors in it's management, but yes, it does feel like the idpol witch hunts have died down a bit since.

>yes, it does feel like the idpol witch hunts have died down a bit since.

You don't seriously believe that was because of the Twitter purchase and not due to a coinciding social shift, do you?

I think the causation is multivariate, but at least partially because of it, yes.

After the Musk acquisition, the identity politics ultras quickly moved to mastodon and bluesky. Regular people were slower to move if at all because of switching friction.

Social media is an amplifier: most of the users are passive consumers or retweeters, and only a smaller minority drive the conversation.

Traditional media also loves to get stories from social media in order to "get the pulse" of society and because journalists are lazy and desperate for narratives. A relatively small social movement can therefore use social media to catapult its message onto traditional media and thus onto society at large.

By isolating themselves from the amplifier, the identity politics ultras were no longer able to push their political message (and witch hunts) to such a large audience, the message stayed confined in spaces that were basically invisible to the average person, and the message faded away.

Yes it wasn't just because of the Musk purchase, but it was definitely a factor.

Token usage is a different and more sympathetic heuristic than LOC produced.

The metric by itself tells you nothing about what value those tokens produced, but to some extent it represents the amount of thinking you are able to offload to the computer for you.

Wide breadth problems seem to scale well with usage, like scanning millions of LOC of code for vulnerabilities, such as the recent claude mythos results.

The trouble with rewarding token usage is the same as rewarding LOC written/generated - if that's what you are asking for then that is what you will get. Asking the AI to "scan the entire codebase for vulnerabilities" would certainly be a good way to climb the leaderboard!

Absolutely, no one should be rewarded for either tokens used or LOC generated. I just think in the absence of any incentives, token usage is a better heuristic as to value produced than LOC generated.

Fantasizing about having someone fired, making no effort to try and understand the viewpoint of the object of one's contempt, does not seem empathetic to me.

As a side note, I found it ironic, that Keith's email that Bryan linked to making the argument that "Empathy is a core engineering value", uses the word "retarded", which by 2013 was already something you could get "cancelled"(or at least chastised) for, because it's not empathetic to the mentally disabled.

Twitter does have a significant amount of racist content, the antisemitism comment is interesting, because while it does exist, at least in my experience twitter seems to be the most supportive platform on Jewish and Israeli issues at the moment.

> [video] It's not free speech

It is though.

Of course it's your choice if you want to post your content there or not, but objectionable speech, _is_ free speech, and if you believe in free speech, then you should protect the speech that you don't like, because one day someone might decide they don't like your speech, and you won't be able to object to it without being admonished for the obvious hypocrisy.

You're taking me slightly out of context there, but my intent was: it's not about free speech. That is, I strongly support free speech (they have the right to be as racist as they want!), but that isn't what this is about: this is about consequences of deplorable (but non-criminal!) behavior -- and just as people have the right to be hateful, we have the freedom to not want to be associated with the racist biker bar that is what Twitter has become.