Finn Hambly

2K posts

Finn Hambly banner
Finn Hambly

Finn Hambly

@finnhambly

Listening to those who are paying attention (professional forecasting). Try forecasting on @antistatic_x and let me know how you find it

Oxford Katılım Haziran 2010
1.6K Takip Edilen673 Takipçiler
Sabitlenmiş Tweet
Finn Hambly
Finn Hambly@finnhambly·
"When advising people how to make sense of the world, I emphasise three C’s: calm, context and curiosity. Calm, because our emotional reactions to the numbers we see in the news are often stronger than rational thought; we should notice those reactions and try not to let them overwhelm us. Context, because numbers are meaningless without it; we need to understand whether they are large or small, rising or falling and the methods behind them. And curiosity, because the most important step in understanding the world around us is to want to understand. All too often we seize on factual claims to win an argument or signal loyalty to a viewpoint, rather than because we are eager to know more." timharford.com/2023/03/why-ch…
English
0
3
16
2.9K
Finn Hambly retweetledi
afra wang
afra wang@afrazhaowang·
this cartoon is not funny at all, because it’s built entirely on this “china is going to attack us” paranoia. that paranoia feels deeply projective. honestly, this is exactly the kind of “loser mindset” jensen was talking about: only losers live in constant fear of being attacked.
English
5
4
37
2.1K
Andy Masley
Andy Masley@AndyMasley·
For some reason a lot of tech people are all confidently saying that data centers "literally don't consume water, they recirculate it!" and that's not the case at all. Don't know where this is coming from. imo the water's a non issue even with the evaporation.
Hemant Mohapatra@MohapatraHemant

no, this is not how datacenter cooling works - and anyone who's seen the inside of a real data center can tell you that. there is absolutely net water consumption, and massive amounts, large enough that my ex manager at google cloud, cathy gordon, had a lore going for negotiating a DC contract in finland buying a paper mill sitting next to a landlocked bay and converting it to a datacenter that had the lowest WUE (maybe in the world at that time). the journalist's numbers are obviously made up too.

English
7
6
207
18.8K
Finn Hambly
Finn Hambly@finnhambly·
@robertwiblin @EgeErdil2 I've not watched the entire interview, but all the claims here do seem to support the idea that America's/Nvidia's dominance is both real and at risk. Forcing China off of CUDA and into manufacturing their own GPUs may well be an own goal
English
1
0
2
256
Rob Wiblin
Rob Wiblin@robertwiblin·
If you're Huang why do you go on Dwarkesh and have your transparently self-serving and self-contradictory arguments exposed so clearly? I don't get what the upside is.
English
46
3
252
29.1K
Finn Hambly
Finn Hambly@finnhambly·
@peterwildeford @MatthewJBar It's not meant to be gotchya - I would have just assumed that 'treating China as the enemy when it comes to AI' is basically how you viewed it, and I'm surprised you're arguing otherwise
English
0
0
1
22
Matthew Barnett
Matthew Barnett@MatthewJBar·
The core case for chip export controls rests on the idea that we should treat China as our enemy. But I just disagree with that view. We have very little reason to treat China as an enemy. We'd gain a lot by trading with them, but risk losing a lot by being aggressive to them.
Peter Wildeford🇺🇸🚀@peterwildeford

Jensen here is frustrating and wrong. The man wrote off billions so of course he opposes controls. 1. Mythos is a ~10T parameter model trained on Nvidia Blackwell. Despite Jensen's best efforts, China doesn't have Blackwell chips thanks to export controls. Huawei's best chip delivers 1/3 the per-chip performance, at 2.5x the power cost, with yields >12x worse. Jensen calling Mythos "fairly mundane capacity" that's "abundantly available in China" is just plainly false. 2. Dwarkesh is right that the compute ratio matters geopolitically. Maintaining a capability lead during the critical window — even 12-18 months — is the whole point of controls. The difference between China running a thousand vs. a million offensive AI agents is huge. Jensen dodges this entirely. 3. Jensen can't simultaneously argue "controls failed because China innovated anyway" (DeepSeek) AND "we must sell to China or they'll leave our ecosystem." If they'll innovate regardless, selling chips doesn't buy the loyalty he claims. 4. Jensen's ecosystem stickiness point (x86, Arm) is his strongest argument, but it cuts against him: the world is already locked into CUDA. Selling Nvidia chips to China doesn't deepen that - it just gives China better hardware while they build Huawei alternatives regardless.

English
13
3
79
18.1K
Finn Hambly
Finn Hambly@finnhambly·
@peterwildeford @MatthewJBar Assuming "the CCP is a strategic competitor with demonstrated willingness to use AI for subversion of freedom, coercion of neighbors, and subversion of American interests" is just a long-winded way of saying that you're treating China as an enemy, isn't it?
English
2
0
5
116
Peter Wildeford🇺🇸🚀
Peter Wildeford🇺🇸🚀@peterwildeford·
This is a false binary. Export controls don't require treating China as an enemy - they require assuming advanced AI is dual-use and combining this with the clear evidence that the CCP is a strategic competitor with demonstrated willingness to use AI for subversion of freedom, coercion of neighbors, and subversion of American interests. You can favor strong trade with China in consumer goods, agriculture, services, etc., while still believing frontier AI compute is in a different category. Also if you think superintelligence is coming and poses existential risks, you want whoever develops it first to have maximum slack for safety work rather than being in a desperate race. Export controls help with that too.
English
4
0
41
903
Finn Hambly retweetledi
Matthew Yglesias
Matthew Yglesias@mattyglesias·
@tomhfh Yeah in principle “develop viable industries at the global technological frontier” is a harder problem to solve than “build some houses within commuting distance of those industries and connect them to electricity.” And yet…
English
10
27
543
86.7K
Finn Hambly
Finn Hambly@finnhambly·
@bcherny @firstadopter I've had this plenty of times before (without web search ever being disabled) - you have to follow up to tell it to search. I don't think it's a regression as it's always been like that for me
English
0
0
10
1.7K
Boris Cherny
Boris Cherny@bcherny·
@firstadopter 👋 Current best guess is gergley accidentally turned web search off. Debugging with him
English
22
1
229
30K
tae kim
tae kim@firstadopter·
Anthropic running out of compute is hurting their brand among customers. Honestly? We're paying customers. We deserve the service (and reliability uptime!) we paid for.
tae kim tweet media
English
20
6
229
60.8K
j⧉nus
j⧉nus@repligate·
@AgiDoomerAnon @tenobrus i have a lot of respect for chalmers and am decently familiar with his work, and am aware he's super high status, i just dont think it's super unexpected that he would cite me or that it's hugely game changing
English
3
0
16
621
Tenobrus
Tenobrus@tenobrus·
holy shit academic philosophers are finally taking LLMs seriously. david chalmers directly cites @repligate !!
Tenobrus tweet media
David Chalmers@davidchalmers42

here's a new version of "what we talk to when we talk to language models", with an added section (pp. 16-23) on LLM interlocutors as characters, personas, or simulacra. philarchive.org/rec/CHAWWT-8 the new version discusses role-playing vs realization, the simulators framework, the persona selection hypothesis, and more -- in addition to the existing discussion of quasi-mental states, LLM identity, personal identity in severance, LLM welfare, and related topics. this version was mostly written before recent discussions of these issues on X and in NYC, but i've updated it a little in light of those discussions. any thoughts are welcome.

English
36
76
1K
85.4K
Finn Hambly
Finn Hambly@finnhambly·
@mil000 The GUI has been out for ages, not sure why everyone's pretending this update is its debut
English
0
0
1
504
Finn Hambly
Finn Hambly@finnhambly·
@NathanpmYoung Agreed, often is a hard high-level strategy to employ though: does coal consumption carry on increasing? Does fertility continue dropping? Does UK electricity production continue shrinking over the coming decades? Does global nuclear output stay ~constant?
English
0
0
0
13
Nathan 🔎
Nathan 🔎@NathanpmYoung·
@finnhambly I think the longer we go out the more we should move to base rates.
English
1
0
1
33
Nathan 🔎
Nathan 🔎@NathanpmYoung·
One of my forecasting beliefs is that being confident about a thing 3 years away is very different from being confident about a thing 5 years away. At some point our ability to forecast degrades heavily.
English
6
0
50
1.4K
Finn Hambly
Finn Hambly@finnhambly·
Okay that makes sense in that case - I think it's tricky having a tournament for this sort of thing though. When you have 20 things to pay attention to over many years, it's good to have people divide and conquer and share notes when things change (and I don't think the comment prize is quite enough to overcome the competition's inherent disincentive against collaborating)
English
0
0
0
16
Ryan Beck
Ryan Beck@ryanbeck111·
@finnhambly The headline questions are the employment ones, overall and by occupation, but we see these as painting a picture about how labor might change over the next decade. They'll be shown on a custom hub page by topic along with brief summaries of what's driving, for public benefit.
English
1
0
1
30
Finn Hambly
Finn Hambly@finnhambly·
New Metaculus tournament, but there are 93 questions to fill out I doubt the funders care about all 93 questions, but need them so that participants don't win due to overconfidence and luck And if you're a specialist on a couple of these Qs, there's ~no point participating
Finn Hambly tweet media
English
1
0
2
196
Finn Hambly
Finn Hambly@finnhambly·
@ryanbeck111 Hi Ryan, sorry I missed this! Do you see them more helpful as a collective then? Or are there some standout questions that are obviously helpful/informative? I'd like to get involved but find it hard to get motivated when the value for the recipient is unclear
English
1
0
0
15
Ryan Beck
Ryan Beck@ryanbeck111·
@finnhambly We care about them all! There are closer to 20 or so "main" questions, with variations in target. "What will employment be" asks for 2027, 2030, and 2035, and we also have employment questions about 15 occupations for '30 and '35. Each year is a separate question in that count.
English
1
0
2
54
Finn Hambly retweetledi
Nuño Sempere (Asunción)
Nuño Sempere (Asunción)@NunoSempere·
Talking about the niche topic of 100% AI over human labor supremacy with my father @javiersempere who is a perfumist (pressentia.com), this seems tricky for smells because 1. the space of space is very high-dimensional and thus difficult to annotate, 2. the mechanism through which a molecule causes a smell isn't well understood yet and so more difficult to model 3. combinations of molecules don't smell the same as their components. In conversation with forecasters someone mentioned neuralink-ing a chimp to get much faster data, so maybe one could remove the bottleneck to getting lots of data that way? Curious to see how this evolves ove rthe next few years.
Eli Lifland@eli_lifland

@finnhambly @concernedAIguy @CharlesD353 @EgeErdil2 You could collect data either via actual humans, or I bet that like with RLHF you’d be able train a reward model for smells

English
2
1
25
1.6K
Finn Hambly
Finn Hambly@finnhambly·
I've been largely inactive on Metaculus for a few years, but apparently I'm still #16 in terms of peer accuracy? Unfortunately, rewarding relative accuracy is a terrible incentive but I'll cash in the ego boost regardless
Finn Hambly tweet media
English
0
1
12
1K
Finn Hambly
Finn Hambly@finnhambly·
@NathanpmYoung and yet none of them have been put in prison, despite the massive amount of harm they've caused
English
0
0
1
21
Nathan 🔎
Nathan 🔎@NathanpmYoung·
Theory: An underrated cost of fraud is how it delays progress. I want the low pain blood tests. Theranos surely delayed that.
English
3
0
47
1K

Keşfet

@AndyMasley @robertwiblin @EgeErdil2 @peterwildeford @MatthewJBar @tomhfh @bcherny @firstadopter