gmm1990

Yeah if you give the llm the exact same input over and over you really shouldn’t be using an llm.

It’s interesting the author ponders what they might be missing now. it seems like it’s the deterioration of the writing due to llms. I’ve liked some of the authors posts in the past. And might look in more detail at the differences between the old posts and new ones because it could be a great example for what to watch out for when using llms for writing and maybe coding.

There weren’t any concrete examples of bugs besides the Google Docs issue. Like what was so bad about blackboard (I also remember thinking it was terrible). There was a link to a great article about search but again no specific example here. There are a lot of words but not a ton of new information. Maybe I’m just bias against llms, this could have been written entirely without then for all I know.

I hope the test data is the same when comparing the different runs. So the big o notation should be the same across different runs.

It’s probably got phone home a few times to to make sure they’re measuring user engagement

This seems to be a very generic/common response to any ai critique. It kind of reinforces my point there’s a lot of situations where the appropriate harness isn’t some agent that’s set to ultra high thinking mode. Chat mode gives the better response and answers the question more quickly

I find that thinking/agent mode sometimes makes it worse/comes up with the same thing and just takes a long time. But I’m sure it’ll be different with fable for a few months until that hype blows over

I guess I’m coming at it from an optimizing market data provider perspective once you preallocate memory the next thing to optimize is the string decimal conversion if the feed isnt binary encoded

That’s quite a bit slower to process. At least if you’re converting to integers to do the calculations and the calculations would be quite a bit slower if you kept the big decimal type

There’s quite a few historical cases of mergers acquisitions being poorly valued aol time warner for example. So maybe v the calculus is different from the execs perspective but it’s not necessarily any better.

I wouldn’t call comparing yourself to Fabrice Ballard and not just saying he’s a better programmer modest.

I don’t know like chess engines didn’t kill chess. You could just play with people that don’t use the “engine”

Not if it drives up energy prices and makes other businesses that employ more people less competitive. Not saying that is the case but it’s certainly not a given

Some of the utilization comparisons are interesting, but the article says 2 trillion was spent on laying fiber but that seems suspicious.

The whole point is that the outages happened not that the ai code caused them. If ai is so useful/amazing then these outages should be less common not more. It’s obviously not rock solid evidence. Yeah ai could be useful and speed up or even improve a code base but there isn’t any evidence that that’s actually improving anything the only real studies point to imagined productivity improvements

If there is really amazing stuff happening with this technology how did we have two recent major outages that were cause by embarrassing problems? I would guess that at least in the cloud flare instance some of the responsible code was ai generated

I think he could have been instrumental to the iphone (not saying he was or wasn't) and whatever he tries next is a complete flop. The ability to be successful is contextual, and great artists can produce mediocre art.

Qualcomm press release: https://www.qualcomm.com/news/releases/2025/10/qualcomm-unve...

Is there a public generic measure of IT outages with historical data. Severe outages seem to be more common lately, but I don't have any data to back it up.

value doesn't have to be monetary

Interesting that there's only the m5 on the macbook pro. I thought the m4 and m4 pro/max were at the same time on the macbook pro

Thanks for the write up. Seems like a cool pattern I hadn’t heard of before

that is an interesting use case, I hadn't thought about a setup like this with a local redis cache before. Is it the typical advantages of using a db over a filesystem the reason to use redis instead of just reading from memory mapped files?

Strange unit of measurement. Who would find that more useful than expected compute or even just the number of chips.

Ah interesting, I was just curious. I’ve wasted some time setting up ci runners stuff on bare metal servers just because I’ve heard runners from gitlab/github can be expensive

That makes sense. Maybe it’s easier in an organization/ some people’s mental model to put guards around changing database because it’s separate from the code, standard across many organizations and in my opinion just harder to change.

I wouldn’t think gpt5 is any better than the previous chat gpt. I know it’s a silly example but I was trying to trip it up with the 8.6-8.11 and it got it right .49 but then it said the opposite of 8.6 - 8.12 was -.21.

I just don’t see that much of a difference coding either with Claude 4 or Gemini 2.5 pro. Like they’re all fine but the difference isn’t changing anything in what I use them for. Maybe people are having more success with the agent stuff but in my mind it’s not that different than just forking a GitHub repo that already does what you’re “building” with the agent.

I seem to gravitate towards nosql type databases, defining tables in a ddl and then again in the code seems repetitive, and slows down changes. But the idea would be that the code is what defines the table. It'd be nice though to hear some of the drawbacks of this. Maybe for very relational things it makes sense to be able to write join queries so data is completely repeated, but my understanding would be that most data base engines would already compress that repeated info pretty well.

Why not run some open source ci locally or the google equivalent ec2, if you’re already going to the trouble of this much customization with running GitHub ci?

How are there not agents that are "instruct trained" differently. Is this behavior in the fundamental model? From my limited knowledge I'd think it'd be more from those post model training steps, but there are so many people who don't like that I'd figure there be an interface that doesn't talk like that.

Extremely anecdotal but all I keep seeing is relatively stable services (the google one comes to mind) having major outages. I assume its not AI related or directly ai related at least, but you'd think these outages would be less common if AI was adding so much value.