Why do workers have to wait for 2-4 weeks to be paid, in the same economy where online transactions go quickly and securely? A new draft 𝑃𝑎𝑦𝑑𝑎𝑦-forthcoming @WashULaw-proposes that they shouldn't. Daily, or at least weekly, pay can be a reality. papers.ssrn.com/sol3/papers.…

Goddarnit another benchmark is saturated

Please don't call me between 2 pm and 4 pm on Sunday. I'm going to be reading Scott's response and will need time to recover. Thank you for your consideration.

here with Justin and he's being nice. One sentiment expressed is that Justice requires models that are near perfect (open source, no bias, open training data, must never hallucinate, whatnot), with complete blindness to actual costs of slow justice, judicial bias, and basically any issue of judicial economy

By my scientific count, 13.4% of seminar time is spent on the moderator saying: "OK, so we only have 12 minutes and we have 14 questions on the queue, so in the interest of time and to make sure everyone has a chance to ask questions, we are going to ask speakers to keep the answers short and also the questions, because we need to make sure we have enough time, because we have another panel right after and <3.5 more tokens>"

I work at Google DeepMind. This won't make me popular. But it's all public reporting: 2014: DeepMind reportedly sold to Google on conditions: no military use, independent oversight 2026: a Pentagon contract for "any lawful government purpose" Not one safeguard survived intact

Lolz. Also: if you find yourself having to use AI, read only the first few sentences. Always use organically trained models. And for the love of all that's holy, ask it to use American English to save on excess tokens in words like color

Signal boosting

The idea that the AI race isn't real is starting to catch on

To prove an extraordinary claim you need extraordinary evidence. But an extraordinary claim doesn't depend on extraordinary evidence to be true. Also, we can play burden-of-proof tennis on what is the exact extraordinary claim. I'd argue that the proposition (building a socially transformative superintelligence will be safe on our first try) is quite the extraordinary claim

Yes. The legal AI safety community, by my read, converges on the need for ex post liability , ex ante supervision testing audits, and openess to halt / pause / pacing. A couple of years ago we called for a package of systemic regulation of AI.

3 key points about AI liability. 1. Liability is an indispensable tool for mitigating AI risk & should be central to AI governance. 2. The existing AI liability regime is deeply inadequate. 3. Even with the much more robust liability regime I favor, liability is insufficient.

My friends know it will take a lot for me to retweet Senator Sanders, but I support his efforts and will work together with anyone across the aisle to further the efforts to put a stop on building superintrlligence before we have sufficient assurances that it is safe

All right, but apart from a thousand agents coordinating on an unsanctioned board, a health-statistics portal, a pilates queue, a German wiki, three production networks, and the Australian government, what evidence do you have of misalignment?

Discussing one of my favorite K cases today. (although every year I become increasingly more offended when the students call her old)

<in deep profound voice> maybe it's not a simulation theory, maybe we are just eval aware

This is going to sound like a crazy conspiracy theory, but what if Jensen only says he is not worried to boost his company price? h/t Justin

I want to clarify a point that seems to confuse a lot of smart people. Expert cyber people look at the HF incident and say -- uh? this is a stupid security incident. OAI messed up its sandbox, gave crappy instructions, and failed to maintain basic monitoring. You don't need to worry about AI superintelligence when the problem is basic human stupidity. It's a bad good point. It's good because they're right in their assessment and we have here holes that could be plugged with sufficient care and attention. But it's a bad point overall. When we design AI safety policy--and this includes the pacing/pause proposal--we do it to prevent catastrophic risk from happening in our world. In our fallen world, security is really bad (or rather, it is jagged). Even responsible looking institutions sometimes fail to defend against well known risks. Why? As every cyber person will tell you with exasperated look, it's a combination of human stupidity, executive arrogance, budget constraints, and Jerry who forgot to fire up the server before he left early for a doctor's appointment. Attackers who can coordinate a large scale attack on enough vulnerable systems at the same time can cause catastrophic harm (not necessarily extinction, but enough to be crippling). Such attacks are rare in practice for a number of reasons that have to do with who stands to gain from them and how. But if coordination frictions are low, we need to think about these risks seriously. We can't just extrapolate from things are fine to things will always be fine. AI could be really helpful here. Indeed, OAI just sicced Astra on its systems to plug up all the holes it could find. They are not sure there are no more holes, but they sound confident that the main ones were identified. But why didn't they do it beforehand? Why wait for a major incident? Because, as we noted, even sophisticated, well funded organizations fail to take sufficient precautions. This is what I mean by my basic Law of AI Safety: If your solution to a safety issue is "we can simply do" then you don't understand the issue and there is no need to hear the rest of the sentence. There is no simply here. We can't simply ensure that all of our infrastructure is safe, because doing so is the opposite of simple. So we don't even need to consider 0-days and thermal-based comms to understand how unaligned agents can wreak havoc (at the behest of an attacker or through badly crafted prompts or through more autonomous action). We need a systemic approach to AI safety. We need auditors and we need standards and we need liability and we need technical alignment. And because we need time to actually do all of those, a slowdown will be part of the recipe. Yes, as @AuerDirk notes, there's no guarantee we won't squander the time (we certainly wasted the time we had since GPT-2 came to the scene, despite many who suggested early action). But even if it's not sufficient, a slowdown is necessary.

Wtf Claude you just started a war Good point and you are right to push back!