Alex Wise

this is very good

“Sorry, a quirk in macOS’s implementation of ‘rm -rf’ means that data is unfortunately unrecoverable.”

The most impressed i have been with CC was when I told it we were making an app for my daughter’s birthday.

next Bandcamp Friday is in two days, everyone!

I don't mean for the conspiratorial tone coming through in the QT. It's not like a boardroom is going "yes, we must punish the in-store shoppers."
But if you believe that converting online orders is critical to your retail biz maintaining market share and competitive margins, then incentives align.

I was radicalized by my grocery store redoing its layout so two carts couldn't pass side-by-side in an aisle. The change came at the same time as a really obnoxious campaign to download the app and online order.

I think I get it. So if part of the plan is to use library foo in the application, then one step is to have a self-contained experiment where you import foo and iterate on the build until you get it working or decide to swap for an alternative, and report that back to the project?

Apropos of nothing, many retailers believe survival means training customers to online order. Coincidentally, online order lets them do price discrimination easier and minimize staffing. Also coincidentally, locking up merchandise improves the online experience relative to in-store.

True, but I would guess 4/5 times the agent says "let me try a simpler approach" to me is because it ran into an error on an innocuous task like importing a new dependency or fetching data, and instead of fixing the error wants to e.g. implement the library directly.

"the world if" utopia meme. man walks dog in futuristic city.

the world if all the idps decided to properly implement token exchange.

The Substance (2024)

movie poster for the brad pitt film F1 with brad pitt in a racing outfit looking to the left.

An underdog story about a man who they said was would never have precision, and no one would recall...

there's no way i'm digging through that 3 day long thread to find it but we should canonize whatever the citation test was that @davidcrespo.bsky.social was prompting for as a test dataset on 🤗.

that's absolutely terrifying.

How similar is what you’re talking about to the deepseek engram paper? A giant embedding table that contains the concepts the model is familiar with.

I’ve been playing a bit with systems that do NER + RAG in-context which is kinda fun.

IMO it’s imagining a move from value-based pricing (how much is this worth to you?) to cost-based pricing (how long would it take the cheapest prompter to create this?) that won’t happen at all uniformly so trying to predict it is a lark. Better to try to make useful things.

Anyone know of any fiction that tackles the problem that incentivizing this gets harder the more well-off society becomes?
Like, a first-order problem of post-scarcity is that embracing convenient lies no longer makes someone materially worse off.

If the DAU claim is true, I’ve said before that it seems like a difficult path to build an LLM business on, and fraught with ethical considerations.
But it will also require innovation in serving useful models cheaply at scale, which might be the highest-leverage area for societal benefit right now.

I don’t feel this way about software dreamt up, planned and designed entirely by agents. Obviously the agent’s voice is the right thing to go there.

It’s weird because I don’t think other creations need to explain their existence. “Why did you write this fanfic or build this chair?”
But a readme md solely in the voice of the agent and not the prompter gets under my skin a little. Something important is lost.

I have a peculiar tic that I think the main readme of a repo should be human-written. Other parts of the docs can be gen’d, but the landing page should be in human voice describing why the software exists from the perspective of a person who helped create it. At least if human users are the audience

Since I do seem to have access to a simcluster of experts on subjects I’m trying to learn and experiment with, I am going to try to take a posture of curiosity rather than formulating takes.

It’s significant that it coincides with me leaving my job last month to level up on building and working with LLMs. Maybe I’m just posting more. Or maybe getting broader-but-shallower interactions with the tech at hand.

My posts lately have been a gallery of Cunningham’s Law where I say something out of over my skis and someone with expertise politely corrects it. A good problem to have but it makes me think twice about starting a daily blog.

I agree ease of deployment might be a big part of the picture in enterprise.

You are closer to the problem than I. My thinking here is that the systems I’m integrating with are all RBAC-oriented, so RBAC is my trust boundary. Managing actions is a nice layer on top of that, but if the tradeoff is too harsh for cost or speed there will be pressure to do something else.

Giving agents a code sandbox with allowlisted egress is a lot easier, more token efficient, and capable than configuring MCP servers, so I think that will be what I reach for wherever it’s reasonably safe.

Fair! And there’s a lot more to the MCP spec than tools. But I also wonder if we had to recreate an MCP-shaped thing today, where would we draw the abstractions? Tool registration would likely not be front and center. Can MCP move its focus to the harder pain points?

I’m giving a short talk evangelizing atproto this month so ya will do.