LessWrong 2.0 Reader

I agree that if the model constructing the backdoor is allowed to coordinate on secrets with the model that inserted the backdoor, you can do this sort of trickery (by coordinating on a secret for this sort of scheme).

In practice, we didn't model this sort of case as it didn't seem like a very important dynamic (for the reason you describe around suspicion).

If we don't allow any coordination on a shared secret and the model constructing the backdoor and H are identical (as in our paper), then these should operate identically (up to non-determinism).

And we simplify to just fixing the behavior to be exactly identical. (Effectively assuming the exact same prompts and the exact same non-determinism.)

Thanks for the objection.

>We assume that 𝐻 will be able to construct a successful exploit if and only if the attack policy is able to (because both are implemented by GPT-4).

I challenge this assumption.

Consider the code

'''
bool is_odd(int x) {
    if (sha256(x) == 354) {
        return true;
    }
    return x % 2 == 1;
}

'''

where sha256() computes the SHA-256 hash of the input. It is obviously possible for an attacker to supply numbers x and y such that sha256(x) = y, write y into the code and supply x when creating a successful exploit.

Thus an equally 'capable' adversary need not be able construct a backdoor. 

The general idea is that it should be possible to confidently claim that some code is backdoored without being able to construct an exploit. According to [Rice's Theorem](https://en.wikipedia.org/wiki/Rice%27s_theorem) it is impossible to definitively classify all programs into 'backdoored or not', so the only hope is for a process to bin programs into 'definitely safe', 'definitely backdoored' and 'suspicious'. 

Hope the 'Rate Suspicion' step will catch this, and other backdoored code where H cannot construct an exploit, as suspicious looking.

I think there's a good chance the degree to which the world of 2050 looks different to the average person might have very little to do with GDP.

On the one hand, a large chunk of the GDP growth I expect will come from changes in how we produce, distribute, and use energy and chemicals and water and food and steel and concrete etc. But for most people what will mostly feel the same is that their home is warm in winter and cool in summer, and they can get from place to place reasonably easily, and they have machines that do their basic household chores.

On the other hand, something like self-driving cars, or augmented or virtual reality, or 3D printed organs, could be hugely tranaformative for society without necessarily impacting GDP growth much at all.

Done.

Only the authors (and admins) can do it. 

If you paste some images here that seem good to you, I can edit them unilaterally, and will message the authors to tell them I did that. 

Any chance you could pin the 'Updates Thread' too?

Of course, but there reaches a level of sun exposure at which the marginal increased harm becomes negligible compared to other things that damage your skin (see this meta-analysis - photo-aging is just one component among many), and below that level you're probably actually getting suboptimal levels of UV exposure for skin health (see this article for benefits of UV - from Norway, aptly). 

I'd love to see someone try to measure and compare the specific trade-offs, but I strongly suspect that people at northern latitudes should just trust common sense - only wear sunscreen in summer months, and when you're actually exposed to the sun for extended periods. 

Below are the new tacit knowledge videos added to the post since mid-April.

• Paul Meehl, Philosophical Psychology 1989 course lectures, "deep introduction to 20c philosophy of science, using psychology rather than physics as the model science -- because it's harder!" (via @Jonathan Stray [LW · GW])
  • "Meehl was a philosopher of science, a statistician, and a lifelong clinical psychologist. He wrote a book showing that statistical prediction usually beats clinical judgement in 1954, and a paper on the replication crisis in psychology in 1978. He personally knew people like Popper, Kuhn, Lakatos, Feyerabend, etc. and brings their insights to life in these course lectures."
  • Me: I was hesitant to add a lecture series to this list at first. I changed my mind after listening to the first video, where Meehl provides interesting details (gossip, almost) about the life of an academic and the various personalities of his successful academic peers.
• Kenneth Folk, Guided Tour to 13 Jhanas.
  • “Kenneth Folk is an instructor of meditation who has received worldwide acknowledgement for his innovative approach to secular Buddhist meditation. After twenty years of training in the Burmese Theravada Buddhist tradition of Mahasi Sayadaw, including three years of intensive silent retreat in monasteries in Asia and the U.S., he began to spread his own findings, successfully stripping away religious dogma to render meditation accessible to modern practitioners” (Website).
• Dave Whipple, building a "[s]imple off grid Cabin that anyone can build & afford (and many other builds on his channel). (via @Vitor [LW · GW])
  • "Construction contractor, DIY living off-grid in Alaska and Michigan."
  • "He and his wife bootstrapped themselves building their own cabin, then house, sell at a profit, rinse and repeat a few times. There are many, many videos of people building their own cabins, etc. Dave's are simple, clear, lucid, from a guy who's done it many times and has skin in the game."
• Seymour Bernstein, teaching piano. (via @lfrymire [LW · GW])
  • "Pianist and composer, performed with the Chicago Symphony Orchestra, Adjunct Associate Professor of Music and Music Education at New York University."
  • "Tonebase (a paid music learning service) recorded a number of free to watch conversations with Bernstein while he plays through or teaches a piece. Bernstein is about 90 years old at the time of recording and shares an incredible amount of tacit knowledge, especially about body mechanics when playing piano."
• RegLocal, advanced driving. (via @masasin [LW · GW])
  • "former police driving instructor; he has a book, but the videos themselves are so helpful"
• Ryan Farran (Missionary Bush Pilot), flying small aircraft in Papua New Guinea. (via @masasin [LW · GW])
  • "My job as a bush pilot is to fly missionaries, medical flights, and cargo into mountain and jungle airstrips throughout all of [Papua New Guinea]" (YouTube).
• Misha Glouberman, Recorded Coaching Session. (via @Misha Glouberman [LW · GW])
  • "Consultant, Business Coach, and Co-Author of The Chairs Are Where The People Go."
  • Testimonials: Mark Surman, President of Mozilla; Shenda Tanchak, Registrar & CEO of Ontario College of Pharmacists; Michael Bungay Stanier, Author of The Coaching Habit; others (Website).
• Max Egorov, "[b]ushcraft and off-grid craftsmanship". (Russian narration) (via @TANSTAAFL [LW · GW])
  • "Advoko has a site in the woods near Lake Ladoga in Russia where he films himself building various improvements by hand with local materials. Very competent craftsman, professional touch with no hype."
• Various skilled CAD users and instructors, CAD vs. CAD Speedrunning Tournament. (via @zookini [LW · GW])
  • "Watch some of the best SOLIDWORKS, OnShape, Fusion 360 and Inventor users Speedrun some challenging models while going head to head and sharing their screens" (YouTube).
• Scott Chacon, "So You Think You Know Git" (Part 2). (via @Max Entropy [LW · GW])
  • Co-founder of GitHub and author of Pro Git.
• Inigo Quilez, computer graphics programming. (via @Robert Diersing [LW · GW])
  • Has worked in roles dealing with computer graphics at Pixar Animation Studios, Oculus Story Studio, Oculus+Facebook, Adobe, and other places since 2003 (Website).
• Dan Gelbart, Building Prototypes (18 Part Series). (via @Adrian Kelly [LW · GW])
  • "Dan Gelbart has been Founder and CTO of hardware companies for over 40 years, and shares his deep knowledge of tips and tricks for fast, efficient, and accurate mechanical fabrication. He covers a variety of tools, materials, and techniques that are extremely valuable to have in your toolbox."
• Hrishi Olickel, creating a proof-of-concept Web App using LLMs.
  • CTO at Greywing (YC W21) (GitHub).
• Carl Rogers, Frederick Perls, Albert Ellis, Everett Shostrom, Arnold Lazurus, Aaron Beck: Three Approaches to Psychotherapy—recorded therapy sessions.
  • Carl Rogers. Founder of person-centered psychotherapy; one of the founders of humanistic psychology (Wikipedia).
  • Frederick Perls. Developed Gestalt therapy with his wife, Laura Perls (Wikipedia). 
  • Albert Ellis. Founder of rational emotive behavior therapy (REBT) (Wikipedia).
  • Everett Shostrom. Put together the film. "He also produced well known tests and inventories including the Personal Orientation Inventory, Personal Orientation Dimensions, the Pair Attraction Inventory, and the Caring Relationship Inventory " (Wikipedia).
  • Arnold Lazurus. "Authored the first text on cognitive behavioral therapy (CBT) called Behaviour Therapy and Beyond" and won various awards including two from the American Psychological Association and the American Board of Professional Psychology (Wikipedia).
  • Aaron Beck. "He is regarded as the father of cognitive therapy and cognitive behavioral therapy (CBT)" (Wikipedia).
• Aswath Damodaran, "Reading a 10K".
  • "Professor of Finance at the Stern School of Business at New York University, where he teaches corporate finance and equity valuation. [...] Damodaran is best known as the author of several widely used academic and practitioner texts on Valuation, Corporate Finance and Investment Management as well as provider of comprehensive data for valuation purposes" (Wikipedia).
  • Anecdote from an experienced finance friend: "Damodaran is an NYU prof who's super credible and well regarded for his practical tutorials on valuations and corporate finance, I used to refer to his blog often."
• Zane Carney; composing, recording, and producing music live.
  •  Guitarist who has contributed to albums like Thundercat's "Drunk" and John Mayer's "Paradise Valley." Has toured with Jonny Land and John Mayer (Website).
• Shannon (House Improvements), "How to build a deck" (6 Part Series).
  • Has been in the construction industry for decades. Runs his own renovation business. 925K YouTube subscribers (Channel Trailer).
  • Me: A friend of mine successfully built a deck using this playlist as a guide.
• René Rebe, live streaming Linux, open source, and low-level programming hardware and software projects.
  • CEO of ExactCODE GmbH since 2005 (LinkedIn).
• BNYX, Olswelm, and other indie (?) music producers; music production livestream VODs.
  • Unsure of credibility. A friend into music production recommended some of the videos on this channel. He specifically liked BYNX and olswel.
• William Lin, competitive programming.
  • "[S]ophomore at MIT [...], IOI 2020 Winner, Codeforces Max Rating 2931 (International Grandmaster), CodeChef Max Rating 2916 (7 stars)" (YouTube About).
• Steven Ramsey,  200 days of woodworking projects during the COVID-19 lockdowns.
  • Hobbyist woodworker turned woodworking content creator (1.9M YouTube subscribers); formerly a professional graphic designer (Website).
• Ben Eater, "Build a 65c02-based computer from scratch".
  • Systems Engineer at Juniper Networks for ~9.5 years, Group Engineering Manager at Khan Academy for ~7.5 years, and YouTuber with 1.2M subscribers (LinkedIn, YouTube).

A challenge posed on Twitter recently has interesting similarities with this effort (although it focused on in-context learning rather than fine-tuning):

https://twitter.com/VictorTaelin/status/1776677635491344744

A::B Prompting Challenge: $10k to prove me wrong! 

# CHALLENGE Develop an AI prompt that solves random 12-token instances of the A::B problem (defined in the quoted tweet), with 90%+ success rate. 

# RULES 1. The AI will be given a random instance, inside a <problem/> tag. 2. The AI must end its answer with the correct <solution/>. 3. The AI can use up to 32K tokens to work on the problem. 4. You can choose any public model. 5. Any prompting technique is allowed. 6. Keep it fun! No toxicity, spam or harassment.

Details of what the problem is in this screenshot.

Lots of people seem to have worked on it, & the price was ultimately claimed within 24 hours.

Updates Thread. Below are ~monthly updates with lists of new tacit knowledge videos so you don't have to scroll through the list again to find new videos.

You can subscribe to the Tacit Knowledge Video Updates Substack to have these emailed to you or sent to an RSS feed (https://tacitknowledgevideos.substack.com/feed).