Shape the Technology that will Shape your Future

    Nagymaros AI Retreat

    17 - 25 August 2022

  • Topics

    Large Language Models

    We are going to be playing with OpenAI's GPT-3, a massive language model trained on hundreds of millions of documents crawled from the Internet. GPT-3 is trained on the mundane task of predicting what the next character is likely to be in a sentence. It can generate random text and complete stories in a remarkably human way. It can create bedtime stories, write essays, learn the meaning of made-up concepts, even solve basic mathematics problems.

    Crazy Image Manipulation

    We're also going to play with DALL-E 2, a relatively new AI method that can generate photorealistic images as well as abstract illustrations or art, based on text prompts. Unlike previous methods, DALL-E 2 has a remarkable ability to "imagine" things that one would never see in the wild: a grizzly bear trying on shirts at H&M.

    What can possibly go wrong?

    AI capabilities develop at amazing speed. New models and breakthroughs in performance are announced almost on a monthly basis. If you're just starting university now, who knows what this technology will look like by the time you graduate. We will also talk a bit about AI safety, alignment and ethics. The field of AI safety looks at ways things can go seriously wrong as this technology gets more powerful. AI alignment tries to develop techniques and processes to ensure AI it's used for more good than bad and that it follows instructions in ways we intended it to. Meanwhile, AI Ethics is broadly concerned with ensuring that (dumb) AI that is already deployed everywhere can be trusted to effect society in a fair way and with respect to people's privacy.

  • Agenda

    Wednesday, 17 August

    • arrival at Dézsma utca
    • 18:00 Welcome and Dinner

    Thursday, 18 August - kick-off workshop @OOOffice

    • 9:00: introductions
    • talk on large language models
    • talk on CLIP and DALL-E 2
    • 12:30 - lunch at Piknik Manufaktúra
    • talk on AI Safety and Ethics
    • project selection

    19 - 24 August - project work

    • 9:00 breakfast
    • generally working a lot
    • also boardgames, walks, sports, with bonus fireworks on 20 August

    Thursday, 25 August - demo day and workshop

    • Project presentations and staying goodbye
  • Example Projects

    This is what participants did during the retreat in 2022

    Political compass

    What does OpenAI GPT know about the policy position of political figures or parties? Can it accurately place politicians on the political spectrum? In this project, Bogi and Eszter designed prompts for GPT, e.g. “Would Barack Obama agree or disagree with the following statement” and then one of the 30+ political statements from the political compass online questionnaire. The responses were then automatically analysed. The project is going to be extended to see if GPT can accurately reproduce expert opinion from the Chapel Hill Expert Survey.

    Codenames: Language model-based word association game agent

    Hédi and Júlia created a system of prompts for GPT so it can play the word association card/board game called Codenames. In this work, GPT plays the role of the code master, as well as the ‘teammate’ who has to guess cards given the clue word. The GPT-based team does OK out of the box, but it starts to struggle with the full 25-card version of the game. The solution is fine tuning on successful rollouts.

    Can GPT learn to play chess?

    Béla and Csaba explored wether GPT can play chess competently using test encoding of games. Overall, while GPT knows many good textbook openings, it struggles to make valid moves after a while. The team explored fine-tuning, and discussed MCTS as possible improvements. While GP understands steps, it doesn’t seem to be able to connect to an ascii encoding of the chess board state.

  • Venue

    Nagymaros, Hungary

    broken image

    Dézsma utca

    Our home for the week is a nicely refurbished Swiss-style villa close to Nagymaros-Visegrád train station. The house has communal areas, a kitchen, three rooms for 3/4 students each as well as additional single and double rooms for others.

    broken image

    OOOffice Nagymaros

    Open-air co-working terrace overlooking the river Danube, on top of popular local food bar Piknik Manufaktúra. The space is fully covered, has a projector for presentations, space for laptops and coffee.

  • Speakers and Mentors

    Among others, the programme and guidance will be provided by...

    broken image

    Ferenc Huszár

    Associate Professor of ML and AI, Cambridge

    I have no idea what I'm doing, but I nevertheless appear confident about it. I studied Computer Engineering at BME (Budapest) then did a PhD in Probabilistic Machine Learning at Cambridge. I worked in tech for 10 years, including startups, venture capital, and four years at Twitter. I joined Cambridge as an Associate Professor in 2020, and haven't done much since then.

    broken image

    Kevin Wang

    AI Research Intern at Redwood Research

    Kevin is a 17yo research intern at Redwood Research, where he’s currently developing mechanistic explanations for how the gpt2-small language model computes behaviors. He likes interpretability, philosophy of science, reading and exploring places. His personal website is https://kevinrowang.com/

    broken image

    Mina Remeli

    PhD student in AI Ethics, MPI Tübingen

    Mina is about to start her PhD with Moritz Hardt and Bob Williamson, leading researchers in fair and responsible machine learning. She studied Computer Engineering at BME, Budapest and received a DeepMind Scholarship to complete an MPhil in Advanced Computer Science in Cambridge. She is broadly interested in attacking, poisoning, breaking, generally messing with machine learning models.

    broken image

    Nitarshan Rajkumar

    PhD Student in AI, Cambridge

    Nitarshan is a second year PhD student in Cambridge, previously a student at MILA, a top ML lab in Canada. He is interested in large language models, topics of AI safety, and general artificial intelligence. He also has an impressive repertoire of about five things he can say in Hungarian (and maybe even more).

  • Participants

    We'll have a very nice group of people to work with.

    broken image

    Anna Kerekes

    Fourth year Mathematics Student at Cambridge, multiple EGMO medalist.

    broken image

    Anna Mészáros

    About to start a Masters in Mathematics in Cambridge after finishing a BSc in Pure Mathematics at ELTE. Former EGMO medalist.

    broken image

    Szilvi Ujváry

    Imperial College Mathematics graduate, about to start an MPhil in Machine Intelligence in Cambridge.

    broken image

    Zsiga Telek

    Starting third year of the Mathematics Programme at Imperial College. Former junior chess champion.

    broken image

    Bogi Nagy

    About to start grade 12 in Szatmárnémeti. Former EJOI and current EGOI contestant. Did a short project on image classification.

    broken image

    Júlia Éles

    Just finished 12th grade in Szatmárnémeti, about to start CS at ELTE in Budapest. Former EGOI silver medallist.

    broken image

    Eszter Halász

    Competitive programmer who just finished grade 12 in Szatmárnémeti. About to start CS at Babes-Bolyai in Kolozsvár.

    broken image

    Csaba Dékány

    Competitive programmer from Győr, about to start CS at ELTE in Budapest.

    broken image

    Hédi Kovács-Bánhalmi

    About to start 10th grade in Kecskemét, interested in robotics, CS, maths and physics. Member of successful WRO and FLL robotics teams.

    broken image

    Béla Bálint

    About to start 12th grade in Szeged, Béla likes maths and coding, has experience with python and coding projects.