whymauri

Anyone can be a founder.

Not everyone can be a leader.

I'm a firm believer that the best benchmark is playing around with the model for like an hour. On the type of tasks that are relevant to you and your work, of course.

My CS:GO lootbox and weapon skin holdings have outperformed my 'serious' investments.

By like, a lot.

Am I crazy or is this paper incomplete? Section 3, Results -- the tables are missing.

I know the area well since we had a family business around there. It's a few blocks away from a sketchy neighborhood; rapidly gentrifying today, though.

At one point it was the intersection of 3 sketchy neighborhoods, but it's been a while.

I'd say there's a selection bias during big international events. You feel it more than say, NYC or Chicago, because baseline ridership is low in Miami. So when Ultra comes to town there's a larger effect size, IMO.

Inspections for aerospace parts are, in theory, a bit more involved than just 'looking at it.'

I watched this video recently and found it really informative w.r.t. commercial aviation near misses:

https://www.youtube.com/watch?v=mIiPt1YVkP8

It gives just enough detail around the classification system, rate of incidents, root causes, etc, that I could search for and learn more on my own after.

I worked at a startup where the CEO wanted each technical contributor to write a blog post per month, every month.

I wrote two internal memos and then quit. Exactly zero got made public. Imagine thinking this is a good use of time for a startup right after closing Series A.

One per IC per month.

I'm confused, I used (what I thought was) DirectFile for Massachusetts via MassTaxConnect (totally free). I make more than that threshold.

Did I do something wrong?

Edit: I understand now. DirectFile is for Federal and MassTax is for State taxes.

Here's an explanation:

https://www.levels.fyi/blog/openai-compensation.html

Yes and no. In transient oriented development you can build the transit infrastructure first and then densify around the new nodes of transit.

I've seen generative models for molecular structures produce results that looked non-sensical at first glance; however, when passed along to more experienced medicinal chemists they identified a bit 'creativity' that only a very advanced practitioner would understand or appreciate. Those hypotheses, which would not be produced by most experts, served as an anchor for further exploration of novel structures and ideas.

So in a way, what you say is already possible. Just how GMs in chess specialize in certain openings or play styles, master chemists have pre-existing biases that can affect their designs; algorithms can have different biases which push exploration to interesting places. Once you have a good latent representation of relevant chemical space, so you can optimize for this sort of creativity (a practical but boring example is to push generation outside of patent space).

From https://huggingface.co/gradientai/Llama-3-8B-Instruct-Gradie...

>We build on top of the EasyContext Blockwise RingAttention library [3] to scalably and efficiently train on contexts up to 1048k tokens on Crusoe Energy high performance L40S cluster.

>Notably, we layered parallelism on top of Ring Attention with a custom network topology to better leverage large GPU clusters in the face of network bottlenecks from passing many KV blocks between devices. This gave us a 33x speedup in model training (compare 524k and 1048k to 65k and 262k in the table below).

My understanding is you need multiple GPUs to coordinate ring-attention for the long context window.

password managers are growing, but I'm not sure that 'most' people use them. Maybe 'most' software engineers or techies, but the average person probably has no idea what a password manager is.

You'd think, but VCs are lemmings and the moment you convince one to invest a clock starts ticking that makes thorough diligence difficult for everyone else.

Get one person to fall for 'just trust me bro' and the hype train follows.

I think LLMs provide more utility than further dependence on cars.

Maybe I'm too corporate-pilled, but if the 'taped up heuristics' provide noticeably better performance for real-world problems, then I don't really care that there is a facade layer around the model itself. In fact, I would pay for that difference in intentional design/optimization if one vendor does it much better than another for my use case.

Is it possible they are using some sort of specialized prompting for these? I'm not familiar with how prompting optimization might work in LLM benchmarks.

I mean, I use GPT-4 on the daily as part of my work and it reliably delivers useful information. It's actually the exception for me if it provides garbage or incorrect information about code.

It's worth mentioning the altitude. I climb several grades less at ~3000 meters (Colombia) than near sea level (Chattanooga, New England, etc). I cannot imagine what climbing at 6k meters+ feels like. And these guys are hauling gear with no margin for error.

This is a mixed route, so the comment above cites three grading scales:

https://en.wikipedia.org/wiki/Grade_(climbing)

The Mixed Climbing, Alpine Ice, and Aid Climbing grade scales, respectively. I wish I could provide more detail into each than the Wikipedia article, but the reality is each grading scale could probably warrant a small book of history and local/regional ethics.

The Internet is dying.

As someone who has worked on ADMET risk for algorithmically designed drugs, this is a nothing burger.

"Potentially lethal molecules" is a far cry away from "molecule that can be formulated and widely distributed to a lethal effect." It is as detached as "potentially promising early stage treatment" is from "manufactured and patented cure."

I would argue the Verge's framing is worse. "Potentially lethal molecule" captures _every_ feasible molecule, given that anyone who has worked on ADMET is aware of the age-old adage: the dose makeths the poison. At a sufficiently high dose, virtually any output from an algorithmic drug design algorithm, be it combinatorial or 'AI', will be lethal.

Would a traditional, non-neural net algorithm produce virtually the same results given the same objective function and apriori knowledge of toxic drug examples? Absolutely. You don't need a DNN for that, we've had the technology since the 90s.

It's an arms race. To feel safe, you need a larger car. When you get a larger car, now other people feel unsafe. And so the bloat goes, until the best selling cars in the US are glorified tanks and Ford doesn't even manufacture sedans anymore.

They are dangerous to pretty much anyone and anything not inside of them. Only 15% are actually used for work.

These lists exist but they are usually not in English. There's countless memes or listicles of Colombian departments on Spanish-speaking websites.

For some reason, searching for conversions can fetch the wrong result for me. But writing the conversion in the search in for Google Chrome (and waiting for the response) is always correct.

Also relevant: https://news.ycombinator.com/item?id=27695574

>Miyazaki would love Stable Diffusion.

You don't know that.