simsla

Any papers you'd recommend?

The thing that bothers me about most paywalls is that they want you to take a subscription to remove it. I've got more than enough subscriptions in my life.

Why does no one use the Google Pay or Apple Pay integration for single articles? Have me pay a buck for an article, and I'd happily do so.

Automatic Phoneme Recognition (APR). There are some models that do this, but they're only so-so.

It looks more similar to the 1929 crash to me, where "too big to fail" blue chip stocks were overinvested and overvalued, and the value adjustments rippled through the rest of the economy. If NVIDIA does get a meaningful value adjustment downwards, it'll probably survive, but it'll impact the S&P500. People will need to sell off other stocks to cover the losses, etc. etc.

There's a pretty big difference between "we need laws and regulations" and "let Trump do whatever he feels like today."

How does a billion break down into 49 million? I'm assuming you mean 400 and 90 million?

100%. It really depends on the application domain and system complexity.

I see a lot of people speaking about a 10x productivity improvement and so on. When I work on hobby projects, I do see that. Just last weekend, I set up a hobby project that I've been thinking about for a while. I'm pretty sure it would have taken me at least a week to implement manually, but instead, it took me three hours.

But some of the systems I have to work on during the day are so big and complicated that you can spend multiple days on a small feature or even just tracking down a bug, even with the support of Opus.

Expecting even a 2x productivity improvement on some those systems is wildly unrealistic. I'm seeing a lot of people get stressed out because the productivity gains from simple application building trickle into the expectations for these complex systems.

That said, if things keep improving at this rate it might just be a matter of time.

I set up mine with WSL and used a Linux terminal over xserver. Once you have a decent unix shell (and a good terminal) it's fine. VSCode etc work fine. This was 2020ish, things might've gotten easier now.

But I do prefer a Linux or Mac for development, just because it's so much less hassle to set up.

Windows developers who don't set up a good terminal environment... I honestly don't know how they manage.

I think there are no safe harbor investments at this time. Even gold is unpredictable.

Personally I went 80% world excl US and 20% equal weight S&P500 to hedge against what I think is an AI bubble. But if the market decides to adjust Nvidia's valuation 20% downward next week, I expect there to be ripple effects throughout the economy.

(Like the .com bubble, I think the tech is genuinely transformative and here to stay, but the valuations are just ridiculous.)

The notation is supposed to mean: you have a matrix Q, and also a shared K=V matrix.

I agree with GP that it's super confusing to us the minus sign as a delimiter between formulas. The tuple notation suggested elsewhere would be way clearer.

IMO the easiest way to open the straight is for the US to back down. "The EU helping the US" doesn't sound like a great way to achieve that.

I don't want the EU to join a protracted Iran war.

I agree with the sentiment, but this is not something I can send my colleagues.

You can be annoyed and right, and still avoid being crass.

Established accounts are worth money, often for scamming/propaganda.

Not too dissimilar to people bot-leveling in MMOs to the sell the accounts.

I'm afraid you're mistaken. He hired a cartographer to iterate over the design, but from the images, he likely used that feedback to create a map style.

For example https://wiki.openstreetmap.org/wiki/Map_style

However, you're also kind of correct in that the rendered tiles are typically cached server-side (presumably also for Apple Maps).

Genuinely curious: what would you have had him do instead?

Everyone is writing. Nobody is reading.

I do wonder if it's fair to expect users to absorb cache miss costs when using Claude Code given how untransparent these are.

Where I live, your employer basically has to give you notice (weeks to months, depending where you live). It's common for that notice period to turn into "garden leave" though, i.e. get paid but don't show up.

Mass layoffs, or RIFs, operate under slightly different rules, but I still saw a stark difference between US and EU employees when I went through one at a different corp.

US accounts were deactivated same day. EU employees were given until end of week to look over the proposed terms etc.

Typical stages of training for these models are:

Foundational:

- Pretraining - Mid/post-training (SFT) - RLHF or alignment post-training (RL)

And sometimes...

- Some more customer-specific fine-tuning.

Note that any supervised fine-tuning following the Pretraining stage is just swapping the dataset and maybe tweaking some of the optimiser settings. Presumably they're talking about this kind of pre-RL fine-tuning instead of post-RL fine-tuning, and not about swapping out the Pretraining stage entirely.

I think my experience as an interviewer has helped. If you ask non-leading questions, sycophancy doesn't come into play as much.

Instead of saying "are you sure?" or "shouldn't we do X instead?" you could say "give me the benefits and drawbacks of this compared to X".

Also, when you yourself are sure, give clear stear. "This overcomplicates A, let's do B instead."

For reference, I knew a guy who built bespoke cheats (less likely to get caught by ban waves) and he charged a few thousand per project.

I've never experienced this, but I guess I always respond with something like "No, [critique/steer]" or "Mostly fine, but [critique/steer]".

Choose to enable backups.

The blog post literally explains how to do so.

Also because normal usage has predictable usage patterns, which allows them to optimise and predict costs. Flat rate pricing only makes sense in that regime.

While I agree, if you need high profits to survive, you're not off to a great start as a nonprofit.

There is a financial incentive to make the search results worse. (More searches, more ads, more money.)

There is no incentive for adding false positives to lists of malicious websites.

Permissions scoping

I've used stuff like this for a hobby project where "effort to write it" vs "times I'm going to use it" is heavily skewed [0]. For production use cases, I can only see it being worth it for things that require using an ML model anyway, like "summarize this document".

[0] e.g. something like the below which I expect to use maybe a dozen times total.

Main routine: In folder X are a bunch of ROM files (iso, bin, etc) and a JSON file with game metadata for each. Look for missing entries, and call [subroutine] once per file (can be called in parallel). When done, summarise the results (successes/failures) based on the now updated metadata.

Subroutine: (...) update XYZ, use metacritic to find metadata, fall back to Google.

And that's probably the opposite of being school-shaped. School is hellishly boring if you're ahead of the curve.