tarvaina

I came here to quote the same sentence. Here's another way to look at it:

Suppose there actually is a best word choice. The LLM doesn't know what it is but makes a guess. Maybe it's the best one, maybe it isn't. The probability that SynthID changes the best choice to a worse one is equal to the probability that it changes a worse choice to the best one.

The main linked article actually does not have that video; the article linked from in the description does have it. Not surprising that someone missed it.

I live nearby and really enjoy Oodi. It's usually much busier than in these pictures.

Nitpick: The early moves in the game are called fuseki. Joseki refers to well-studied local patterns of moves and they appear through the middle game, not just in the early game.

A couple of things I love about go is that you don't need to memorize fuseki, and that applying joseki correctly is as much a matter of judgment as it is of memory.

(I am a 1 dan go player but haven't played much in the last 15 years.)

The Wikipedia article says:

"Rock, sand and concrete has a heat capacity about one third of water's. On the other hand, concrete can be heated to much higher temperatures (1200 °C) by for example electrical heating and therefore has a much higher overall volumetric capacity."

and

"Polar Night Energy installed a thermal battery in Finland that stores heat in a mass of sand. It was expected to reduce carbon emissions from the local heating network by as much as 70%. It is about 42 ft (13 m) tall and 50 ft (15 m) wide. It can store 100 MWh, with a round trip efficiency of 90%. Temperatures reach 1,112 ºF (600 ºC). The heat transfer medium is air, which can reach temperatures of 752 ºF (400 ºC) – can produce steam for industrial processes, or it can supply district heating using a heat exchanger."

(Off topic.)

Not only that, but Linus's parents were politically active communists and young Linus was a pioneer (like a boy scout but for communists). His father also lived in Moscow for several years on two separate occasions.

_Miles_ from the ocean? Budapest is 620 miles from the nearest ocean at the mouth of Elbe.

(Incidentally exactly 1000 km.)

You are absolutely right, I cannot see it there either! Sorry for the misdirection.

You can turn them on in the workspace admin settings: https://help.openai.com/en/articles/11954883-legacy-model-ac...

It took me a while but I think the difference between Vertex and Gemini APIs is that Vertex is meant for existing GCP users and Gemini API for everyone else. If you are already using GCP then Vertex API works like everything else there. If you are not, then Gemini API is much easier. But they really should spell it out, currently it's really confusing.

Also they should make it clearer which SDKs, documents, pricing, SLAs etc apply to each. I still get confused when I google up some detail and end up reading the wrong document.

Yeah. Alternative explanation: I'd say 3 and 7 are out because they often come up in fairy takes etc as magical. 5 is out because it's half of ten and the number of fingers. 2 is out because it is too small.

Oh, right! I forgot about it. I guess 8 would do too.

Also it is one party (The Finns) presenting a rail initiative competing with their government partner's (National Coalition) older initiative. It is very unlikely that they both will be implemented.

Stray thought: Why 4 and 9? Because the joke is funniest if the number is completely ordinary.

0 and 1 are special and so are all prime numbers. 6 is out because it's the maximum die throw. And one figure is more ordinary than two figures, or negatives, or decimals. That leaves 4 and 9.

The article is about choosing 12 colors from the 4096-color palette.

There are actually six different TV series:

1959: https://en.wikipedia.org/wiki/Die_Muminfamilie

1969: https://en.wikipedia.org/wiki/Moomin_(1969_TV_series)

1972: https://en.wikipedia.org/wiki/New_Moomin

1977: https://en.wikipedia.org/wiki/The_Moomins_(TV_series)

1990: https://en.wikipedia.org/wiki/Moomin_(1990_TV_series)

2019: https://en.wikipedia.org/wiki/Moominvalley_(TV_series)

This is how Clojure does it.

https://clojure.org/guides/learn/hashed_colls#_creating_a_li...

Website of the actual event: https://www.the-final-experiment.com/

Isn't white vs blue collar a latent variable? You have to operationalize it somehow. If you just ask "how blue collar are you?", people's answers will be influenced by all kinds of subjective biases.

I'd argue sitter vs stander distinction also makes this presentation more visceral, memorable and understandable. Collar color would feel unnecessarily abstract and boring.

As a Finnish speaker I've often wondered about this. In Finnish (and Hungarian) the stress is almost always on the first syllable, so it is easy for speakers to distinguish word boundaries. How do speakers of other languages do it?

Now I'm wondering if the first consonant is lengthened in Finnish too or if the syllabic stress makes it unnecessary.

Did I get this right?

Bard – old name of their generative AI service, to be called Gemini

Duet AI – old name for their generative AI in Google Workspace, to be called Gemini

Gemini – three things: 1. the name of their models (like GPT). 2. the new name of their free service (like ChatGPT), gives access to Pro 1.0 but not Ultra 1.0. 3. the new name of the Generative AI tools in Google Workspace.

Gemini Advanced – the name of their paid service (like ChatGPT premium), gives access to both Pro 1.0 and Ultra 1.0

Ultra 1.0 – the first version of their big model (like GPT-4)

Pro 1.0 – the first version of their smaller model (like GPT-3.5)

Google One AI Premium – the subscription that you need to buy to have access to Gemini Advanced

Google One Premium – the old version of the subscription, does not include access to Gemini Advanced

Google app – the mobile phone app, which includes either Gemini or Gemini Advanced

Google Assistant – like Siri but hard to define what it is

Google AI – a generic name for all their AI products

We got access soon after the API was announced and have happily been using Claude Instant in production for a couple of months now. It may have helped that our use case was a good match for their capabilities.

Suppose we have trained a model to perform a certain set of tasks. Later we would want to teach it a new task. Catastrophic forgetting means that teaching it a new task makes it unlearn some or all of its earlier tasks.

It occurs because training changes the weights of the model. The earlier set of weights was good for the previous tasks. The new set of weights is only good for the new task. Usually special care must be taken to overcome catastrophic forgetting.

For questions such as the difference between typhoon and hurricane, ChatGPT and Claude give accurate answers in just a couple of seconds. It’s such a different experience compared to the search engines.

The impossibility of a general lossless compression algorithm is a consequence of the pigeon-hole principle.

https://en.m.wikipedia.org/wiki/Pigeonhole_principle

I would guess max_tokens refers to the number of generated tokens, not the size of the context window. That at least is the meaning of max_tokens in the ChatGPT API.

https://platform.openai.com/docs/api-reference/chat/create#c...

Based on the NPR article you linked this sounds like a normal scientific debate, though? One researcher says that data supports their theory, another says that it doesn't if you look at the data more carefully. I don't see accusations of fraud here, just different interpretation of facts, and a valid challenge to an earlier theory. This is how science is supposed to work, isn't it?

Gino, on the other hand, is accused of outright fraud. Data Colada claims she took real data and changed it to support her theory, or possibly created entirely fake data.

The ingredients you need for training a useful machine learning model are expressivity, learnability, and generalization. Many methods are universal approximators but that only takes care of the first ingredient. Arguably the reason neural networks are so successful is that they can offer a good balance between the three.

Before transformers we built different neural network architectures for each domain. These architectures offered better inductive biases for their respective domains and thus traded off some of the expressivity for better learnability and generalization.

Nowadays the best architectures seem to be merging towards transformers. They appear to offer more generally useful inductive biases and thus a better trade-off between the three ingredients than the earlier architectures.

Tensorflow is also from Google Brain, not DeepMind.

Not all transformers have separate encoders and decoders. GPTs, for instance, only have the equivalents of decoder layers of the original transformer paper, but they are still considered transformers. Karpathy’s video shows an actual GPT-style transformer.

I think a neural network can be considered a transformer if it contains a stack of attention blocks as its core mechanism.