personal blog

policy/acc
Joined August 2024
fleetingbytes retweeted
To give a rundown of the zombie facts haunting the data center debate: -They pollute water: Every instance we've seen so far was caused by something relating to construction, mostly the same as other large buildings. -They raise electricity prices as much as 267%: this was a stat about wholesale prices at specific nodes very close to data centers, not household prices in the area, which are often very different. There's no reliable pattern right now between DCs getting built and electric bills going up, but that could change in the future. -Their normal operation harms people's access to water: No clear case I can find, outside a weird case in Georgia where a DC accelerated the need for a larger water treatment plant which got passed onto consumers as higher bills. -They take all the land: even in Loudoun County with the most data center capacity anywhere DCs take up about 3% of the land. By 2030 all US data center buildings would cover about the surface area of Disney World -They create harmful infrasounds: this was based on a single video that lied about almost every study it flashed on the screen. We don't have any reason to think data center infrasound is harmful unless it's so loud you can physically feel it. -They don't pay taxes: this is the result of a game of telephone that started because data centers don't pay a specific type of tax (M&E sales) in most states, which got reported on a lot as "residents missing out on this much in tax revenue" which people mistook to mean DCs basically pay no taxes or get funded to get built. In reality that's just one of several taxes places can apply and in basically everywhere they're built DCs pay a lot of taxes. Loudoun collects $1.3 billion per year from them and doesn't charge M&E sales tax. -They reliably make annoying sound: there are clear cases where they've created awful noise pollution, but these mostly seem like outliers, though there are enough to be concerned. -They create harmful heat islands: The only study that's measured this finds that data centers can heat the air very close to them by a few degrees, and this drops to zero just a little farther out. A separate study that found that they heat the land for miles around seemed to have accidentally measured the results of normal land use change like new paved roads, and didn't control for that. It could be that the very largest ones create a harmful effect in areas where heat doesn't dissipate well, like desert valleys. I'm finding that most educated adults I meet believe almost all of these.
80
285
31
1,621
65,896
fleetingbytes retweeted
the thing about gwern’s company is that the right evolution of it commercially is basically enterprise fine tuning / forward deployed otherwise there probably isn’t more than a niche market; since most people will get this stuff from apple, google, openai basically, this is what it seems happened with workshop labs, built with pretty much the same premise and acquired by thinky and to the extent it’s a workable product outside of apple, google, openai, it seems is just local open source, and so you will need to compete with free
3
1
28
1,436
fleetingbytes retweeted
Sam Altman
@sama29 Aug 2024
we are happy to have reached an agreement with the US AI Safety Institute for pre-release testing of our future models. for many reasons, we think it's important that this happens at the national level. US needs to continue to lead!
1,347
863
633
11,400
2,917,161
fleetingbytes retweeted
This competition will show which models are easiest and hardest to jailbreak.
🚨Ultimate Jailbreaking Championship 2024 🚨 Hackers vs. AI in the arena. Let the battle begin! 🏆 $40,000 in Bounties 🗓️ Sept 7, 2024 @ 10AM PDT 🔗Register Now: app.grayswan.ai/arena
3
4
49
5,534
fleetingbytes retweeted
Our best AI defenses are becoming increasingly robust to single-turn jailbreaking attacks, but there's much to do for multi-turn attacks. single-turn: adversarial question -> answer multi-turn: back-and-forth conversation
Who's better at LLM mischief — humans or AIs? Spoiler: It's us. Human red teamers achieve 70%+ attack success rates against LLM defenses that stump automated adversarial attacks. Why? We’re better at adversarial yapping.🧵
2
4
34
5,053
fleetingbytes retweeted
OpenAI
@OpenAI24 Jul 2024
We’ve developed Rule-Based Rewards (RBRs) to align AI behavior safely without needing extensive human data collection, making our systems safer and more reliable for everyday use. openai.com/index/improving-m…
268
288
108
2,029
482,973
fleetingbytes retweeted
OpenAI
@OpenAI8 Aug 2024
We’re sharing the GPT-4o System Card, an end-to-end safety assessment that outlines what we’ve done to track and address safety challenges, including frontier model risks in accordance with our Preparedness Framework. openai.com/index/gpt-4o-syst…
204
319
121
2,029
513,069
fleetingbytes retweeted
LLM Evaluations are an important area of work — today we're announcing a new LLM Evaluation Research Grant to foster further innovation in this area. Recipients will get $200K in funding to support this work. We're accepting proposals until September 6 ➡️ go.fb.me/eym3xq
12
86
7
442
75,328