fizx

Interest rates have many 2nd order effects, including VCs getting more funding from their LPs, and future-looking stocks becoming a more appealing investment class. Its not just about loans.

heat

> local RPC calls have to serialize protocol buffer messages

They don't have to, they artificially added that constraint to make the benchmarks fairer.

5ms protobuf, 2ms custom, 0.4ms in-process

They're claiming that the runtime will figure it out for you.

It wouldn't be too surprising if this is the sort of thing that an optimizing compiler or query planner could do better than a human.

If not, you're probably at the scale where performance regressions are caught and rolled back at early phases of rollout.

Like most magic, it's either going to make things 100x better or 100x worse, depending on how leaky the abstraction is at its current state of maturity.

Doesn't MCTS imply that you'd have to generate a whole tree of tokens? Instead of maybe a 200 token response, you'd have to generate several thousand tokens as you explore the tree?

What's the difference between a job loss and a productivity gain?

How often did you release?

There are more ways to know things than observation.

This looks like a llama2 finetune, so the dataset (inclusive of llama2) isn't fully open as claimed, and I'd still have to accept the Facebook and possibly OpenAI licenses.

Let alone that clearly the base model was built on non-source-code, so their premise doesn't hold.

Disappointing.

The most likely hypothesis I've seen for Q*:

https://twitter.com/alexgraveley/status/1727777592088867059

My estimate is that a typical staff engineer who'd been at OpenAI for 2+ years could have sold $8 million of stock next month. I'd be pissed too.

There's much I don't love about Rust, but I feel golang could steal the ? operator and keep the spirit of go.

Effectively,

instead of

    result, err := doSomething()
    if err != nil {
        return nil, err
    } 

you'd get the same control flow with

    result := doSomething()?

Do you think you're smarter than the rest of the team combined? You can't make the "correct" decision by yourself. And then everything gets more complicated.

7 9's is 3 seconds of downtime per year. That was never a thing.

Something like https://tilt.dev/ where you spin up a subset of the service graph in a cloud environment that hot-reloads based on local edits.

There's a saying: closer to the ceiling is closer to the door. The higher up you get, the more likely that you're blamable for problems, and the less likely that you are repurposeable for some other project.

What if replica mom was in your life for ten years before someone told you? I suspect you'd have a brief moment of shock and betrayal, and a few days/weeks later continue on as if nothing happened. (at least with respect to love for replica mom--i imagine you'll struggle with other issues forever)

CGI was the OG serverless, because it was a process-per-request model until e.g. fastcgi came along.

Put scripts on an S3 backed filesystem, and as you grow, wrap them in a container, then an auto-scaling group. It might even work today!

I'm surprised I haven't seen more of this from the anti-Kubernetes crowd.

If I'm an old person on fixed income, and my equity doubles from $1M to $2M, causing my property tax to double, I can take out a home equity loan against my house and be not only just fine on payments, but far ahead of where I was otherwise.

This argument was also a lie.

Yup. Also tickles me that I might have had better equipment than this team, though we still had to source sample carriers from McDonalds.

The multi-million-dollar measurement apparatus required us to put samples in disposable tubes that cost about a dollar each. They turned out to have the exact same physical properties as a McDonalds straw, so once a week someone made a food run and grabbed a handful on the way out.

Some of it, yes.

But in practice, you want to fix a bug in chinese tokenization, or OpenAI releases the next version of its embeddings, or you want to add a few synonyms, or change the aggressiveness of the stemmer.

Then you have to rewrite your whole search index, and if its part of your primary db, you're pretty sad.

One of the many problems with search inside Postgres is that search is very CPU-heavy (and starting to become GPU-heavy). Ideally, you'd like to reserve CPU in your database for transactional updates to your core data model.

I've seen a lot of ES and Solr clusters operating at 100% of 10+ nodes during a re-index, or just 30-50% of 10+ nodes during normal operation. The corresponding database would be say an AWS L/XL instance at 50-100GB of data and 30% CPU utilization. Moving all of the search CPU into your primary DB means now you'd have to shard it.

But I love PG extensions for search, recursive joins, vectors, etc on side projects. It can keep things fun and simple.

We ridiculed the first generation of bluetooth headsets, calling Jawbones "douchenozzles".

People will come around sooner or later on some generation of an apple VRset.

It's a random engineering director in the growth dept, likely not on the hook for Spaces. Feels pretty likely to be unrelated.

Ease-of-use for python devs

1.26% yearly inflation based on the march number alone.

There's no more text to train on.

Also, amazingly, transfer learning works better than anyone could have hoped! As an analogy, imagine that you could show that football players learned musical instruments 50% quicker than non-athletes. That would be amazing!

But LLMs can write music and training them on music makes them better writers. It's almost like intelligence & skills are completely general!

It's intentionally using the naive 2^N solution to stress test lots of tiny function calls.

All 3rd party clients shut down is the best case for twitter.

All 3rd party clients migrate their users to Mastodon, while simultaneously solving the UX and approachability problems that have hampered Mastodon, sure seems like the worst case scenario for twitter.

I've often wanted to extend the https://selectorgadget.com algorithm to be able to do find & replace on code ASTs, instead of the DOM tree.

It'd be pretty cool to be able to find-by-example, selecting snippets of code and having the algorithm tell you what they all have in common.