voiper1

I assume they mean when the phone is "turned off" meaning the screen is off, not powered down.

From the article: "... but did not respond to numerous questions about other aspects of the group’s accounting."

It's not "nitpicking" when the reports of where/how much they are saving money -- their purported goal -- is _fundamentally_ inaccurate in many circumstances that can be verified. And even worse for what that implies about everything else that can't be verified.

My query generated 17 steps of research, gathering 74 sources. I picked "Deep Research" from the modes, I almost accidentally picked "reasoning".

Wow, that's so fun. Just upload and generate assets super easily... queue the next one while it's working.

>To work around this, we had to change the assertion logic in some tests to check the name and code, instead of the message. This is also the standard practice for checking error types in Node.js.

Sounds like something they should try upstreaming?

Now they'll need to track all the tests to manually modify/import...

Cute. First thing I did though was to see if moving my mouse or clicking affected it. Doesn't seem to, please add that!

The first question is always: which model? Which fortunately you at least addressed: >free plan of ChatGPT, but I can remember if it was 4o or 4.

Since chatgpt-4o, there has been o1-preview, and o1 (full) is out. They just announced o3 got a 25% on frontiermath which is what this article is a reaction to. So, any tests on 4o are at least TWO (or three) AI releases with new capabilities.

I read both those articles, but I still don't get how to do it. It seems the idea is that more of the embedding is informed by context, but do I _do_ late chunking?

My best guess so far is that somehow I embed a long text and then I break up the returned embedding into multiple parts and search each separately? But that doesn't sound right.

I thought embedding large chunks would "dilute" the ideas, since large chunks tend to have multiple disparate ideas?

Does it somehow capture _all_ of the ideas, and querying for a single one would somehow match?

Isn't that the point of breaking down into sentences?

Someone mentioned adding context -- but doesn't it calculate embedding on the whole thing? The API Docs list `input` but no separate `context`. https://docs.voyageai.com/reference/embeddings-api

Related XKCD: https://xkcd.com/1172/

:sigh:

With all the concern over AI, it's being used _against recommendations_ to detect AI usage? [0][1]

So while the concern for using AI is founded, teachers are so mistaken at understanding what it is and the tech around is that they are using AI in areas it's publicly acknowleded it doesn't work. That detracts from any credibility the teachers have about AI usage!

[0] https://openai.com/index/new-ai-classifier-for-indicating-ai... openai pulled their AI classifier [1] https://mitsloanedtech.mit.edu/ai/teach/ai-detectors-dont-wo...

Cool, based on geolocation it gets the vacation days for your country.

> Ideally in a web application, aside from some in-memory caches, no object allocated as part of a request should survive longer than the request itself. Any object that does is probably something that should be eagerly loaded during boot, or some state that is leaking between requests. As such, any object promoted to the old generation during a request cycle is very unlikely to be immortal, so promoting it is wasteful.

So could each request clean up it's own garbage when it finishes, so then they should never need any global garbage collection?

Compared to what?

https://en.wikipedia.org/wiki/Cost_of_electricity_by_source#...

Coal is $4000 per kW

This solar project is right on the average at $1,327, although NREL says it can cost upto 2743.

Some investment group says a 1MW coal plan would cost similar or upto 4x the cost: >estimates suggest that the construction cost of a coal-fired thermal power plant can range from less than $1,000 to $4,500 per each kilowatt of installed capacity. https://esfccompany.com/en/articles/thermal-energy/coal-fire...

So, seems just about average and on just the construction cost - cheaper than coal.

Right, I would assume we'd render multiple 2D images from different angles, and train on that.

But as I'm writing this, why can't we do ML directly with the 3D models...

Wow, that's an incredible writeup.

Super surprised that shelling out was nearly as good any any other method.

Why is the average bytes smaller? Shouldn't it be the same size file? And if not, it's a different alorithm so not necessarily better?

Sure, it's more effecient to have it use an API. And people have been integrating those for the last while.

But there's tons of applications that are locked behind a website or deckstop GUI with no API that are accessible via vision.

It's out, and improved!

It went from 77.4% to 84.2%, skipping past O1-preview which is at 79.7%

Source: https://aider.chat/docs/leaderboards/

Is there a tracker that identifies APIs / libraries that do this unicode injection protection?

I'm guessing few do.

I pasted this text into openrouter's playground for 4o, sonnet 3.5, gemini pro and I got a mix of "nothing here" to "it's invisible unicode and I don't process/no standard meaning". So, not so practical already?

Gemini Pro 1.5: >Those characters are from the Unicode Private Use Area and don't have standard meanings. Therefore, they won't display consistently across different systems and fonts. I can't "see" any meaningful text, just a series of boxes or symbols depending on my current setup. It's possible they are being used for some custom purpose within a specific application or system, but without more context, I can't interpret them.

Claude 3.5 Sonnet >The visible text I see in both lines is simply "look" at the end of each line. The rest of the characters before "look" appear to be invisible or non-displaying Unicode characters.

ChatGPT-4o >It appears you're asking what text I can see, but I can only interpret the text that has been input into this conversation.

The human can "audit" the entire text he's copy-pasting and still miss it.

And it provides a way to exfil without it being visible.

Sure. It's worse though, because it provides a way to invisibly in-fil data even with full font sizes, and a way to ex-fil data.

I hadn't seen that one, I love it!

Any recommendations for s3/b2 - anyone can upload (or with password) and only the admin can download?

Goal: allow customers to upload large files.

Aider benchmarks them great for coding... super slow token generation. Much cheaper but once you're used to the speed... it's too slow.

I like this ironic take that even the experts's expectations got upended with new tech.

Article's point seems to be that in the long term: paper count, citations, impact, motivation and fulfillment will all come from focusing on a project.

+1 for Burkerman. The first thing I read from him was "it's worse than you think" [1]

I was trying to finish my to-do lists to rest when he pointed out... I'll never finish. I'll always put new things on my to-do list. Once you dig down, you realize the issue is even worse and that gives the opportunity for a paradigm shift.

https://ckarchive.com/b/75u7h8hkk9g9e

similar to anthropic outage - claude 3.5 is much used by the community for coding.

Sources:

https://openrouter.ai/models/anthropic/claude-3.5-sonnet:bet... - claude-dev: 5 billion tokens this week

https://openrouter.ai/models/anthropic/claude-3.5-sonnet/app... - aider - 325m tokens this week

Also see aider's leaderboard for models: https://aider.chat/docs/leaderboards/

OK I can't even (easily) find the numbers.

What is an average person's carbon footprint? what is a rich person's carbon footprint? 10x, 100x?

Best I see right now is https://www.visualcapitalist.com/co2-emissions-by-income/ which is from 2019 and doesn't say exactly who's covered - I assume it's a global mix?

Income Group Average income (USD) Share of CO₂ Emissions Top 1% $310K 15% Next 9% $90K 34% Middle 40% $16K 43% Bottom 50% $2K 8%

So the top 10% by income account for 49% of the carbon emissions, and the next 40% for 43%, and the lowest 50% for 8%.

So that would mean if I got this right: Top 1%: 94x / year / person Next 9%: 24x / year / person Middle 40%: 7x / year / person Bottom 50%: 1X/year/person

So the top 1% person uses about 14x more carbon than a middle 40% person. Over one order of magnitude. I actually expected it to be far worse, what with the yachts and private planes.

Any better sources/numbers for this?

I'm still new to LoRA/fine tunes, but: I can't just dump in 1gb of data, correct? I need to structure it in Question/Answer or the like?

So it would seem the cost really becomes converting/curating the data into a usable format first.