upbeat_general

And if there are differences, that is often a good thing! It can mean the author wanted to format something in a particular way that the journal didn't allow.

My uneducated guess is that it’s done to avoid clutter for the rest of the settings and because accessibility is a good justification to outweigh this concern.

This is just not how it works - it is perfectly plausible (in fact the most likely) that pretraining was well finished by the time Fable was released. A bit of extra distillation takes far less compute.

Moreover, it’s very plausible (and expected) to use multiple clusters and GPU types for RL rollouts which could very well not be included in this count.

No part of this pipeline is fixed in stone.

Same way you do for planes, cars, etc. You have long(ish) life-cycles and some pieces that can be independent (e.g., same turbine engine on several plane models).

If I used a racecar to go 25mph in a residential neighborhood, I’d make a similar conclusion.

I think that figure (haven’t verified it but assuming it’s true) isn’t complete. It hides who and where those people are - for example, I imagine art history skews towards higher ranked schools in the first place.

If we’re defining on-prem as fitting in a rack - then every frontier model can be hosted on-prem.

Now this might not be the most cost effective (and may require a bit extra power), but you only need a datacenter for training or cost optimization.

This statement is well known to be incorrect for at least a year.

If by "happened to" you mean pour significant resources for well-over a decade on many different AI research groups then yes, that's accurate. Depending on your definition of AI, it might even be two decades.

In fact, OpenAI was founded largely with the direct goal of preventing Google from being the sole winner in AI...

This isn’t quite accurate. Data weighting is quite important in pretraining.

Sampling exists.

From a user perspective it is nice to know that the person decoding will likely support a given format, both now and in the future.

More use cases for a single popular format makes this more likely.

I really wish that there was an “almost yolo” mode that was permissive but with light restrictions (eg no rm), or even better, a light supervisor model to prevent very dangerous commands but allow everything else.

I can assure you that lacking knowledge in DPO (and especially GRPO it’s just stripped down PPO) is not a dealbreaker.

This seems like the perfect case for legislation that starts out targeting higher volume devices/larger companies and lower over time.

I don’t see why the industry couldn’t move to providing this documentation/full source over a few years.

Not saying it’s for sure the case but it might be that the model gets confused by OOD text from the other model whereas it expects its own text to be online from itself (particularly if the CoT is used as context for later conversations).

A lot of developers (including myself) don’t want to notarize/sign their binaries that they want to run on their own machine(s).

And it was none of the mentioned schools!

Seconding another commenter. I went to a state-school (maybe in the top 5 US state schools) and got an offer without any elite background.

Related: I've been pretty frustrated with cursor that you can't call any APIs directly, and it all has to go through their backend. Along with this, they have essentially refused to fix bugs in their custom API key page for months, forcing you to use their "1st party" tokens.

Here’s the simple reason why this is not the case:

If you provide the 6DoF trajectory (+ gripper joints), a lot of robotics (manipulation in particular) is basically solved. The problem is, we don’t have these good trajectories.

Sure, joint space is commonly used for learned policies, but cartesian space isn’t uncommon either.

IK is really just not a major focus on the learning side of robotics because it’s not the problem. The problem is we don’t know what to do even at the slightly higher level.

Not a war crime, China has plenty of agriculture.

See the Cuban missile crisis. It is an act of war but certainly not as escalatory as direct attack.

The challenge here is not IK. Sure, you don’t need IK if your policy directly operates in joint space but this isn’t the motivation for doing RL.

IK to my knowledge is well known in every setting I am aware of.

Has anyone used Typst for academic conferences/journals?

To me it’s sorta like CSS where the availability of templates is more important than the core language (at least for people lacking design abilities).

It’s less reproducible than docker (assuming the pip usage is correct). Docker specifies a lot of OS properties that UV ignores.

That being said, UV is great.

I tried out uv a bit ago and dropped it. But about two weeks ago, I switched to it and migrated two projects with no issues.

Things like pypi sources per dep are there finally.

I still find rough points (as many others pointed out, especially with non sandboxed installs), that are problematic, but on the whole it’s better than Mamba for my use.

imo gemini-exp-1206 is the best public LLM that exists right now.

Clusters and supercomputers are able to fully control the hardware & OS. environment modules are useful in that setting but do not solve OS differences, not to mention themselves need to be installed/maintained. lots of time developers do not work on your same cluster (say, for any open-source code) and environment differences can be significant.

I regularly use large clusters and even there docker/singularity are very helpful at times (very simple example, glibc requirements).

When you are talking about well-written code that only requires posix (and nothing beyond) and does not interface with hardware, etc. etc. then virtualization seems crazy.

For a start, any suite that takes >X hours on a single node, especially compounded if you have a large team of developers.

> At least as long as you're running your own code on your own hardware

Assuming you keep a consistent env/OS across all nodes you will want to run said code. Which can be difficult, even just between two users on a single node.

Not to mention the fact that a lot of (most?) code needs to (A) interoperate with other people's code and (B) at least sometimes run on other hardware.

I think it’s just super handy to know where you’ve been. Great to be able to find places (“Oh I loved this Thai place I went to 4 years ago in X city”)

Also to be able to find photos in specific locations (if you integrate it with geotagged photos like google does).

It’s one of those things I don’t actively care a ton about in the short term but is useful data to have in the long term if it can persist without a lot of effort.

Edit: I think form is “irrelevant” as long as it’s well supported and can be useable in different services later on. Same thing with notes imo (and why I use markdown). Has enough features but importantly is and will be well supported.