philomath_mn

Reminds me of this quote from Tonya Riley's _The Staff Engineer's Path_:

In The Art of Travel (Vintage), Alain de Botton talks about the frustration of learning new information that doesn’t connect to anything you already know—like the sorts of facts you might pick up while visiting a historic building in a foreign land. He writes about visiting Madrid’s Iglesia de San Francisco el Grande and learning that “the sixteenth-century stalls in the sacristy and chapter house come from the Cartuja de El Paular, the Carthusian monastery near Segovia.” Without a connection back to something he was already familiar with, the description couldn’t spark his excitement or curiosity. The new facts, he wrote, were “as useless and fugitive as necklace beads without a connecting chain.”

Why is that?

AI bills are already trivial compared to human time. I pay for claude max, all I need to do is save an hour a month and I will be breaking even.

most people agree that the output is trite and unpleasant to consume

This is likely a selection bias: you only notice the obviously bad outputs. I have created plenty of outputs myself that are good/passable -- you are likely surrounded by these types of outputs without noticing.

Not a panacea, but can be useful.

Anywhere I can follow your takes on LLM-assisted coding?

ChatGPT is $20 / month?

Under what circumstances would that cost be high? Is OpenAI going to rip off your app? Why would they waste a second on that when there are better models to be built?

The exact same logic applies to any deductible expense, and yet people think they can buy a business vehicle "for free" because it is deductible.

IDK why this is so hard to understand.

By building a good reputation and contact list doing salaried work. Usually you do a good job there, make a bunch of stakeholders happy, and then you have a chance at spinning off on your own.

Best part is that they probably have data to show that all that patience costs the typical passenger mere seconds to a minute on 99% of rides.

This has always bothered me about aggressive or impatient human drivers: they are probably shaving like 30 seconds off of their daily commute while greatly increasing the odds of an incident.

Curious what you are expecting when you say "bottom falls out". Are you expecting significant failures of large-scale systems? Or more a point where people recognize some flaw that you see in LLMs?

While this unfortunate, I am sure I also have single lines in production with greater cost and equivalent value (close to none) -- and I've only worked at small companies. I am sure some of y'all can beat this by ~2 orders of magnitude.

Databricks is happy to have us as a customer.

Any of the leading LLMs could answer this, they just need the right context. So paste in the model cards for the various models you care about, or ask a model with search or "deep research" capability.

That being said, there probably isn't too much alpha in chasing down model tradeoffs outside the top few players (OpenAI, Anthropic, Gemini).

That definitely seems to be the play. Keep funneling in users from Lovable/bolt.new and keep building revenue or hope to be acquired if one of those vibe coding tools gets huge.

I agree: the incentives to use more and more AI are too strong. We're all stuck in some form of the prisoner's dilemma and the odds that nobody will defect are much too low.

So it seems the most rational position is to embrace the tools and try to ride the wave before the gravy-train is over.

This is a bit of a strawman. There are certainly people who claim that you can ask AIs anything but I don't think the parent commenter ever made that claim.

"AI is making incredible progress but still struggles with certain subsets of tasks" is self-consistent position.

That is a sub-token task, something I'd expect current models to struggle with given how they view the world in word / word fragment tokens rather than single characters.

Great breakdown, appreciate it.

I think most of the hype around MCP is just excitement that tool use can actually work and seeing lots of little examples where it does.

Watching Claude build something in Blender was pure magic, even if it is rough around the edges.

Not trying to be smug or blithe, but this is one of the main things I love about living in the suburbs or the country: neighbors can only bother me outside (and they rarely do).

It seems like the possible outcomes are:

(a) nobody uses the language, so a package manager doesn't matter OR

(b) people use the language, they will want to share packages, then a package manager will be bolted on (or many will, see python)

Seems like first-class package manager support (a la Rust) makes the most sense to me.

Attack? Is there any other indication of an attack?

There are many possible root causes which could explain what you observed.

Going from an X1 Carbon to a MBP felt like stepping 10 years into the future. The seamless lid close, battery life, operating temp, build quality and performance were all _huge_ upgrades.

I held out on Mac for 20 yrs, no idea what I was thinking.

Let’s be honest though: there has been an immense amount of propaganda on both sides. How many times has Ukraine reportedly embarrassed Russia, knocked them back on their heels, been within weeks of winning? All with young beautiful people on the front lines singing patriotic songs?

Russia is certainly the aggressor, let’s make no mistake. But I honestly can’t make heads or tails of what is actually happening in Ukraine because of all the universal propaganda

Right, I just don't know where the dividing line would be. What % of a book is acceptable?

Idk, it is pretty good a generating synthetic data and recognizing the different logic branches to exercise. Not perfect, but very helpful.

Wonder how Readwise is managing what they do -- really hope they don't get shut down.

I usually throw a lot of context at it and have it write unit tests in a certain style or implement something (with tests) according to a spec.

But the o3-mini-high results have been just as good.

I am fine with Deep Research taking 5-8 minutes, those are usually "reports" I can read whenever.

I paid for pro to try `o1-pro` and I can't seem to find any use case to justify the insane inference time. `o3-mini-high` seems to do just as well in seconds vs. minutes.

In my experience, LLMs have a hard time processing large amounts of raw data and finding patterns (even simple univariate time series) so I would be surprised if a GPT could find _real_, non-hallucinatory patterns in raw data.

I would agree, but I do think I get a lot out of Readwise which helps me review what I have read.

If this could help me with review or highlight collection then I'd be happy with it.