frabcus

15 ms·

211.
by frabcus 3y ago
I agree it isn't a moat in the business sense - that would be some kind of lock in network effect. e.g. If ChatGPT being popular gives OpenAI enough extra training data, they're locked in forever having the best model, and it is i
212.
by frabcus 3y ago
Personally I think trained models are derived works of all the training data. Just like a translation of a book is a derived works of the original. Or a binary compiled output is a derived works of some source code.
213.
by frabcus 3y ago
Also end users complain - I ran an LLM for language learning with millions of users, and we made extra calls to the OpenAI content moderation API because both we and our users wanted it to stay polite and "safe". Most corporate ap
214.
by frabcus 3y ago
What are some specific tech examples where Gemini won't answer?
215.
by frabcus 3y ago
Can it read information from images, like GPT-4V / ChatGPT+?
216.
by frabcus 3y ago
I guess it's competing with ChatGPT+ at about £16 / month. And you get a bunch of extra storage and Workspace features from the Premium Google One too. If it is actually as good as GPT-4, I can imagine lots of people swithing subs
217.
by frabcus 3y ago
I bought a Nebula lifetime to help support that lifeline here. I funnel money to these creators via Patreon. I too would love other ideas!
218.
by frabcus 3y ago
Why's it FUD? It seems likely to me that GenAI can make fake people, and break current KYC.
219.
by frabcus 3y ago
Google Translate and DeepL are pretty bad at this, that's true. As far as I can tell they translate a sentence at a time. Using an LLM and giving more context, it can infer gender from signals earlier in the text. This based on testing
220.
by frabcus 3y ago
Right but users could be given prominent ways to deal with this - eg tools and buttons to (in bulk at first) unfollow people.
221.
by frabcus 3y ago
I had an iPhone 11 and the Fairphone 5 feels like an upgrade. (Case on the former as it shatters and is expensive to repair and no case on the latter)
222.
by frabcus 3y ago
This happened to me too. Every committed iPhone user told me - you should have used a case! Ummm... What's the point of a phone being beautiful and thin if I then anyway have to make it large with a case? It was a bad piece of design.
223.
by frabcus 3y ago
Really loving my Fairphone 5 - basically smartphones are enough of a commodity now everything feels really high quality and fast physically. The sky blue colour is really nice. AND also it avoids conflict metals, is repairable. Much much be
224.
by frabcus 3y ago
Is Llama Guard https://ai.meta.com/research/publications/llama-guard-llm-ba... basically a shared-weights version of OpenAI's moderation API https://platform.openai.com/docs/api-reference
225.
by frabcus 3y ago
I see this slightly the other way round - the difficulties caused by tokenisation are why it is good at segmentation. Words break and jump around due to it, and more so with typos in the vast amounts of training data. Also regarding backt
226.
by frabcus 3y ago
I guess he'll make the consumer hardware product he's been developing with Jony Ive, and also spin up the chip company he was working on anyway.
227.
by frabcus 3y ago
Does the Microsoft deal let Microsoft continue training from e.g. the GPT-4 weights? I guess at least it gives them access to the OpenAI models to use internally, which they kinda need as their ways of working (Greg especially) will be high
228.
by frabcus 3y ago
"I specifically say I’m in favor of slowing down, which is sort of like pausing except it’s slowing down. If we’re at a speed of 10 right now, a pause is reducing to 0. I think we should aim for a 1-2 instead." - Emmett Shear, Sep
229.
by frabcus 3y ago
This could all be good for Microsoft. Microsoft want OpenAI to do research, and give them the model to run on Azure. They certainly don't need OpenAI competing in consumer products. ChatGPT could have been a Micrsoft product in an on
230.
by frabcus 3y ago
It's a non-profit. It isn't a Silicon Valley company. https://openai.com/our-structure
231.
by frabcus 3y ago
On the contrary, it shows it is a non-profit. People keep denying it is and pretending it is a normal Silicon Valley startup. This demonstrates clearly it is a non-profit.
232.
by frabcus 3y ago
Whatever they've built this year presumably uses all the positive/negative feedback on ChatGPT that they have a year worth of data now... Another examples is the Be My Eyes data - presumably the vision part of GPT-4 was trained on
233.
by frabcus 3y ago
The Atlantic Article makes it pretty clear that the fast growth of the commercial business was giving Ilya too few resources and too little time to do the safety work he wanted to do: https://archive.ph/UjqmQ
234.
by frabcus 3y ago
That is a secondary and huge problem, but the larger initial problem is making sure the AI aligns with values that nearly all humans have (e.g. don't kill all humans)
235.
by frabcus 3y ago
Worse. Which is exactly why superintelligence is scary - it'll make the humans around it go wild for power, and then it will be impossible (by definition) to predict.
236.
by frabcus 3y ago
So can Dario Amodei and Mustafa Suleyman.
237.
by frabcus 3y ago
Obvious wall would be financial - Sam arranged more Microsoft funding, diluting the non-profitness even more, and tried to force board's hand by high cash burn and persuasion. Board had to act fast to fix it. And OpenAI changed enterpr
238.
by frabcus 3y ago
Via his Y Combinator related holdings in companies which use the OpenAI APIs? Or just which are getting capital and exposure due to ChatGPT boosting visibility of AI.
239.
by frabcus 3y ago
He was fired from being Chair of the Board, but the rest of the board left him in his position as an engineer (?) in the company. Then an hour or two later he resigned as an engineer. See: https://news.ycombinator.com/item?i
240.
by frabcus 3y ago
Agreeing to that would be a very good reason for a board to sack him.

More ›