frabcus

12 ms·

181.
by frabcus 2y ago
Meta haven't made an open source model. They have released a binary with a proprietary but relatively liberal license. Binaries are not source and their license isn't free.
182.
by frabcus 2y ago
I'd find knowing what's in the training data hugely valuable - can analyse it to understand and predict capabilities.
183.
by frabcus 2y ago
Interesting thanks! Anonyms technology is presumably what the original article was about? Any idea what a transaction will look like. Who will pay them for this and how? I guess ad networks would pay so they can still attribute and do retai
184.
by frabcus 2y ago
How will it help them earn money? Are Google threatening them with withdrawal of search revenue unless they add the feature? Something else?
185.
by frabcus 2y ago
I'd love a detailed description of the politics and economics of adding this. Why are Firefox doing it, how does it relate to similar features in other browsers? If it leads to direct revenue, how? If there are relationship reasons for
186.
Automating Meridian DSP5000 Speaker Activation for One-Click Audio Bliss(github.com)
1 points by frabcus 2y ago|0 comments
187.
by frabcus 2y ago
They both use deep learning and gradient descent. Also both were made possible by availability of GPUs.
188.
by frabcus 2y ago
My experience with models even about a year ago is that the model has firmly decided what it thinks by the time of the last layer, so the probabilities on that layer aren't very useful. You either get the same (in this case wrong) thin
189.
by frabcus 2y ago
The answer is that it is hugely variable between people! Hurlburt has great research on this using Descriptive Experience Sampling. Some people mainly use images, others mainly speech, others mainly emotion etc. And many more use a varied m
190.
by frabcus 2y ago
It's definitely in the excellent Feynman BBC series "Fun to imagine"
191.
by frabcus 2y ago
Yes - to me the speech feels more like a forcing layer that drives thought. The actual thought is a recurrent neural network underneath. Nearest I have of conscious access to itis non verbal awareness of complex concepts.
192.
by frabcus 2y ago
Quite a lot of people have no inner voice, others no inner imagery, others no inner unsymbolized conceptual thinking (cf all of Hurlburts research). We all use very varied modalities of thought! It's as rich as how different we look or
193.
by frabcus 2y ago
It has never been and still is not a company. It's a non profit. It doesn't act like that any more, which is the point.
194.
by frabcus 2y ago
It's there now! And still 128k context window
195.
by frabcus 2y ago
I can't see any calculator for the audio pricing ( https://openai.com/api/pricing/ ) or document type field in the Chat Completions API ( https://platform.openai.com/docs/api-reference/
196.
by frabcus 2y ago
Sagittal.ai | London, UK | On Site, Full Time | https://sagittal.ai/ We're looking for an experienced Senior Software Engineer with interest and familiarity in Large Language Models, architectures and prompting techniq
197.
by frabcus 2y ago
This sounds like it is fully reproducible: "it also included reproducible training recipes that allow the weights (neural network files) to be replicated, which is unusual for a major tech company so far" Digging into the code, is
198.
by frabcus 2y ago
In tests I did with linguists a year ago, GPT-3.5 was better than DeepL also. Google Translate largely caught up with DeepL a while ago! Specifically, LLMs have a context window larger than a sentence. So can eg infer gender throughout docu
199.
by frabcus 2y ago
The GPT 3.5 API is cheaper than Google Translate. And in our testing (over a year ago) was better for translation. So I assume in that case the energy use is less?
200.
by frabcus 3y ago
I see this as huge Affero (AGPL) praise! It would be great if the license was used more. I'd much rather contribute to something where sharealike extends in some way across networks - software doesn't link in binaries, as the GPL
201.
by frabcus 3y ago
Just a note that it is more like prices are 42% higher - because the 30% is a cut off gross, and 100 / 70 = 1.4287
202.
by frabcus 3y ago
Inflection blog post about their new CEO, and their view of the change, and new API access to their model via Azure: https://inflection.ai/the-new-inflection Also it is worth signing up to and trying out Pi, and talking to
203.
by frabcus 3y ago
I consider the weights a binary program and the source code is the training data. The training algorithm is the compiler. I agree this isn't standard terminology, but it makes the most sense to me in terms of power dynamics and informa
204.
by frabcus 3y ago
An article about this with some more details: https://www.euractiv.com/section/politics/news/albania-to-sp...
205.
by frabcus 3y ago
That looks amazing - the home screen focus reminds me of Niagara on Android (although that is prettier)
206.
by frabcus 3y ago
My favourite extension for customising YouTube right now is Unhook. https://unhook.app/ I hide all recommendations and suggestions, so basically all I can do on YouTube is look at channels I've subscribed to, or search
207.
by frabcus 3y ago
This is my favourite YouTube feature! It makes the home page lovely.
208.
by frabcus 3y ago
Just a note that consciousness isn't always or even mostly a verbal "monologue", assuming you mean that. Apologies if you didn't! A significant percentage of people never have an inner voice at all. And few people use it
209.
by frabcus 3y ago
They don't own 49% of OpenAI. They have capped rights to 49% of OpenAI's profits.
210.
by frabcus 3y ago
I'm not sure. If it fails, then I reckon a binary compiled from source code fails top. There's nothing creative about the act of a compiler, it is automatic, just like the training run of an LLM. And no part of the original source

More ›