andy_ppp

3 ms·

1.
by andy_ppp 5d ago
This is why I was careful to specify tokens per second not time to first token which is the prefill step you’re talking about. Clearly to run these models well you need both but as I said adding more chips or compute units can give you more
2.
by andy_ppp 6d ago
Tokens per second is almost entirely memory bandwidth at inference time, training obviously needs more compute but you can add more chips for that.
3.
by andy_ppp 6d ago
The agents will soon decide all this, not the humans ;-)
4.
by andy_ppp 8d ago
It can go either way, you might get a relief of pressure or you might get an increased destabilisation in the region.
5.
by andy_ppp 8d ago
I agree with this, the rich actually are oddly MORE greedy in Europe and take double the equity for half the money (roughly) so founders get a lot less upside or they move their businesses to the US and get US capital at better rates.
6.
by andy_ppp 8d ago
Why can the Chinese build models and Europe cannot? The algorithms behind this stuff are not that complicated, are they? Is it the cost of energy? The illegality and difficulty of obtaining all the data in the world? Lack of capital to star
7.
by andy_ppp 9d ago
Yes I sometimes think the "don't train on my data" is actually a good signal for "this data/person is probably better to train on because they want to keep something private". The whole copyright system should
8.
by andy_ppp 9d ago
You believe any of these companies care about the law? They care about winning and building the self improving AI as quickly as possible.
9.
by andy_ppp 10d ago
So, during the Great Depression who ended up doing well? What can be applied to today?
10.
by andy_ppp 10d ago
Why? Seems the US doesn't like democracy anymore and the EU for all its faults still does. I can see Japan joining too and potentially South Korea, Philippines, and others who are concerned about the state of the world and want to work
11.
by andy_ppp 10d ago
How funny, they chose very beautiful old architecture and very ugly modern architecture and invented a result - why not choose things like the Bahai temple in Delhi or Fallingwater by Frank Lloyd Wright? I'm sure there are some old ugl
12.
by andy_ppp 11d ago
It’s impossible to know for sure but if they’d stayed private I think they’d be more not less successful, just because startups can move more nimbly. Although we’d need to define success, we’d be here forever…
13.
by andy_ppp 11d ago
People here are assuming that current CEOs are not actively destroying their companies, maybe the agent presented here is actually a very accurate simulation!
14.
by andy_ppp 11d ago
Excellent! Let us know when!
15.
by andy_ppp 11d ago
You know there was a lab that was cutting edge in Europe and it is still there - Deepmind - but the UK let it be sold to the US for a fraction of what it's worth today - $400m. We just allow US companies to asset strip the EU and then
16.
by andy_ppp 11d ago
Who decides on when to press said button?
17.
by andy_ppp 11d ago
Maybe we could connect AI to everything and make ourselves completely vulnerable to and dependent upon it instead?
18.
by andy_ppp 11d ago
It would obviously be opt in but it’s just an idea, it clearly might dissolve on contact with reality!
19.
by andy_ppp 12d ago
“I think they all just added a bunch of similar tasks to their training sets, so we cannot judge true emergent capabilities of the models anymore.” This has been the plan since the start of all this, they regurgitate code in ever better for
20.
by andy_ppp 12d ago
Love the output from Qwen 3.8 it seems very impressive for the cost! Why does Gemini 3.8 flash blur everything? What are Google playing at!
21.
by andy_ppp 12d ago
You could “Pokémon Go” this and allow people to find performances they find interesting on a random street corner anywhere! Great idea!
22.
by andy_ppp 12d ago
I love Nike shoes and brand but they just don’t produce anything since fly knits that I actually like. All I want is a flat shoe with a massive toe box and thin material for the gym with Nike branding - why is it so difficult to make someth
23.
by andy_ppp 13d ago
I wonder if people are getting it to help with Krypyos the final part? Great documentary on this code… https://youtu.be/jVpsLMCIB0Y
24.
by andy_ppp 17d ago
Except Taalas is much faster than GPUs, orders of magnitude so. They aren’t going to get 100x faster at inference any time soon!
25.
by andy_ppp 18d ago
Not sure, the labs will probably just cripple the security features of these models for a while I think and even potentially put back doors into systems for the security services…
26.
by andy_ppp 24d ago
I've never found feature flag systems complex and you can build one yourself in a few functions. The services I have used in the past were extremely simple to integrate so I don't understand the issue either way honestly.
27.
by andy_ppp 27d ago
The referendum was done specifically in two parts - 1. do you want us to try to find a deal with the EU 2. We will thrash out any details, get the best deal we can with the EU and give you another vote. Seems like Iceland didn't want t
28.
by andy_ppp 28d ago
Most people do not enjoy the output of the open models or know where their code is being sent so I’m out. Also I don’t trust Elon anyway, I’m just going to make a VS Code Extension that handles things how I want but I’d like to offer the sa
29.
by andy_ppp 28d ago
Why you would trust Elon Musk to not hack you and steal your model weights is beyond me, but maybe they intend to use that compute just for training, still I’m sure there’s plenty of research there to take too…
30.
by andy_ppp 1mo ago
So did they train the AI to be superior to everyone and make bad decisions and never ask for clarification?

More ›