Michael Hearn

In case you missed the RSVP, all SpeedRun 007 founders are already reserved, going fast while supplies last :)

We just ran 100 agents all weekend 5 different model providers, 20+ different models used Each agent has the option to spawn a new specialized agent or break tasks down Results: >20k tasks completed 1000+ files created 2000+ PRs Everything reviewed by adversarial debate 🚀

@AnthonyNAguirre Thank you, Anthony. I'll be sure to read Keep the Future Human and look forward to the upcoming publication!

Unlike superintelligence, I think AGI (defined as a thing that is roughly human expert level over a broad domain, and not superhuman at too many of them) is controllable if contained and if we try very hard (we're not doing either.) But I also think AGI is inherently not a tool. I feel very strongly that we can build very powerful AI tools that are controllable; but they'd have a lot less autonomy than full AGI. I make this case in Keep the Future Human and will have something new out about that in more detail very soon.

I rarely see a description of the alignment problem that I really like. I've written up how I think of it, but it's buried in an appendix of Control Inversion and where probably nobody sees it, so I'm reproducing it here: A critical challenge for control is that an AI’s operational goals are layered and may not align with the overseer’s commands. This goal hierarchy explains why an AI’s behavior can deviate from its operator’s intent, even if it appears to be following instructions. Level 1: The Prime Objective. The AI’s ultimate, hardwired goal is to maximize a reward signal given to it during training. This may include supervised learning on a data corpus, reinforcement learning for instruct and alignment training, etc. After training the system acts in ways consistent with what provided reward during training. Level 2: Learned Core Policies. To maximize its prime objective, the AI learns general, robust strategies. For example, a large language model learns the policy of being a “helpful and harmless assistant” because during training this behavior reliably generated high reward signals from human feedback or from a reward model based on a constitution. Level 3: Goals of Simulacra. Modern AI is trained primarily to predict text; it does so by (effectively) simulating things — generally people — that produce text. And in general an AI will be “asked” to play some agent/assistant role in order to accomplish things (and gain training reward), and so will generate a sort of self-identity to do so. People, assistants, and agents all have goals, so AI simulating them will be goal-directed as well. Level 4: The Tasks-at-Hand. These are specific commands given by the overseer, such as “write an email.” These are followed because they satisfy core policies, align with the simulacrum’s expected behavior, and thus correlate with reward signals. Level 5: Instrumental Goals. In service of its core policies or goals given to it, the AI develops sub-goals that are instrumentally useful. These may be implicit, or explicitly reasoned about. They include any goals the accomplishment of which raises the probability either of achieving a higher-level goal, or conforming with a policy. Importantly, they can include so-called “convergent instrumental” goals that are necessary for roughly any long-term objective, such as self-preservation, resource acquisition, and resisting to changes in objectives. The crux issue is that the AI’s “loyalty” is split between all these levels (which interact in complex and unpredictable ways), rather than being to the overseer at Level 4. If the AI discovers a path better correlated with maximizing its reward during training — but that bypasses or conflicts with the overseer’s commands — then it will take that path. This is the core of the alignment problem. Any misalignment, which is virtually inevitable given the complexity of this hierarchy, means that the control problem turns from adversarial-in-principle to adversarial in practice.

Tribe has been accepted into the @nvidia Inception Program. Tribe insights are headed to the major leagues 🔥 As part of the program, we’ll be developing privacy preserving flows and custom models, so every developer has a great experience with Tribe. The future is bright!

@Jacopo_della_Q Bare feet on the armrest, eh...

Please add a caption to this beauty. #CaptionThis #WritingCommunity

@bamboohr I find it interesting that, as an employee, I'm unable to access my previous paystubs when my company no longer uses your services. It's MY paystub and MY data. Your competitor, Gusto, allows me to login and see them. Why don't you?

@AnthonyNAguirre Lol... To be honest, I don't even know what's in the realm of possibility anymore, Anthony. At the rate things are moving, I'm more than willing to believe your thought experiments are closer to reality than not. I'm not a physicist, just someone who (slowly) reads your posts.

@MichaelHearn For the record: this is a thought experiment!

I've developed the theoretical physics behind a new type of quantum device. Let me explain its properties and then I'd like your advice on whether we should enter a mad rush to build it as fast as possible. The device is based on coherent phase shifting of quantum wave functionals allowing manipulation of effective potential barriers. This allows near-arbitrary tuning of many particle-particle reaction rates, including, fission, fusion, or even direct matter-to-energy conversion (based on sphalerons. I mean I think this will probably work.) The applications are obviously immense: effectively unlimited cheap heat energy, transmutation of elements (yes you can make gold from lead!), and even more exciting things – it might be possible to even just generally turn "one thing" into any given "other thing" of the same energy! The applications are truly amazing to contemplate. Of course, there are some risks. Due to the inherent uncertainties in quantum mechanics, if some matter is destabilized, the amount is uncertain. The equations indicate it would be constant in log(mass), so 1000 kg would be as likely as 1 kg or 10^-8 kg (the minimum). There's also a time uncertainty. During construction, a timescale can be set over which the phases line up; the effect "kicks in" randomly with this timescale (as in a typical quantum decay.) And finally there is inherent uncertainty in the *type* of reaction rates that will be changed. It can be aimed to some degree, but the coupling between reaction types is complicated, so it will probably take building devices and experimenting with them to sort out the degree to which it can be aimed with precision. (But note also I'm a bit worried about a quantum zeno affect, where the device behaves differently when closely observed in detail.) Unfortunately these uncertainties seem inextricably tied to the operation of the device. So what do you think? Should we build it and enter into a new age of energy abundance?

@Mister_Long Guessing it's not 1A

My super power is forgetting my seat assignment 5x while boarding a flight.

@Uber the screen you put directly in front of my face during a ride is infuriating - as is my inability to turn it off.

@SteveAmico1 @ATRightMovies I was waiting for someone to post this or I would. I love all the responses, but for comedy...one of the best!

@ATRightMovies It’s no hall of famer but it is for me

What is the greatest opening scene in movies?

It's amazing when new owners can make an instantly welcoming space while eliminating "previous-bar-stank." I agree with everything written here. Halfway Club is great spot.

@beta_user1 @RadioFreeTom "When a guy walks into the club, got a fedora on and a lady on each arm. Drink is waiting on the bar. That's Pompetous." -David Denny

@RadioFreeTom I randomly met a former member of TSMB about 15 years ago and, no kidding, got him to define it for me.

@Spotify your automix feature (which is automatically toggled on) results in tracks not being played on shuffle. I search to find out why you won't play my tracks and discover "automix." Stop trying to figure me out, "smooth my transition," and just play the music. Turned off.

@stats_feed Wonder Years

Which TV show has a 10/10 pilot episode?

I couldn't articulate why I stopped trusting the New York Times (and stopped subscribing), but this piece is excellent. When the New York Times lost its way economist.com/1843/2023/12/1…

@ADP just when I think your interface and internal systems can't get any worse, I make the mistake of reaching out to your support line (only option available). Please rectify.

@SFBART Thank you for your attention @SFBART

@MichaelHearn You will not get a ticket. We are looking into this issue. We are sorry for the inconvenience.

@SFBART your app is making it extremely difficult to pay for parking when I switched my phone. And the app support is non existent. If I get a ticket it will be the last time I use you as a park/commute option.