Budapest Institute of Advanced Technologies
===

# 1. Proposal Summary

*-- 2 minute read*

I am proposing a new AI research and education institute in Budapest, Hungary. Its purpose is to expose exceptionally talented students to research on AI, AI ethics, AI risks, accelerating their training and guiding them to careers with maximal longterm impact.

The country has an outstanding track record in selecting and growing mathematical talent from a young age. There are books written about the model of maths education pioneered here, and competitive results speak for themselves (Hungary is #5 on the IMO medals table). However, beyond the focus on competitive Maths and CS, students lack opportunities and mentorship that would guide them to impactful careers. Instead, many end up in low-impact academic disciplines or uninspiring technical jobs.

The opportunity to tap into this underutilised pool of exceptional talent is massive. The institute will operate as an overlay to existing schools and universities. We will proactively identify the best students studying anywhere, support them through scholarships, 1:1 mentoring, a world-class accelerated curriculum, research internships offered from a young age. If successful, we will build a scalable pipeline of AI research talent with priorities around ethics and safety.

This institute could usefully leverage significant amount of funding, and, if the model is proven, could scale beyond Hungary to a network of institutes across Central and Eastern Europe.

## 1.1. Why This Would Succeed?
* **leadership:** project lead by [Ferenc Huszár](https://www.inference.vc/about/) - established ML researcher (currently Cambridge Associate Professor) with significant entrepreneurial and leadership experience (startup with $150m exit, VC and M&A experience, contributions to Twitter's AI strategy).
* **existing groundwork:** builds on years of existing pilot projects and grassroots effort funded by Ferenc's personal [funds which he earned to give](https://twitter.com/fhuszar/status/1341045408726986754) as an AI ethics advisor.
* **leverage:** leverages the existing exceptionally strong STEM education in Hungarian state high-schools, and free undergraduate degree programs which provide strong training in foundational subjects.

# 2. Strategy

*-- 10 minute read*

This section follows Richard Rummelt's ([2011](http://goodbadstrategy.com/about-the-book/)) framework and outlines the three key components of our strategy:
1. a diagnosis that explains the nature of the opportunity
2. a guiding-policy for exploiting the opportunity, and
3. a set of coherent actions designed to carry out guiding-policy


## 2.1. Diagnosis, observations

* **unique special maths secondary schools:** Hungary has 12 secondary schools that offer special mathematics curricula where students study up to 2× the number of weekly Maths lessons vs standard curriculum. These programs were developedd as an educational experiment in the 60s (see e.g. [Connelly Stockton, 2010](https://digitalcommons.sacredheart.edu/cgi/viewcontent.cgi?article=1046), [Vogeli, 2015](https://www.google.co.uk/books/edition/Special_Secondary_Schools_For_The_Mathem/Q1kGCwAAQBAJ)). These schools have produced legendary mathematicians and computer scientists who made significant contributions to graph theory, number theory, probability theory, etc. The schools are clearly outstanding internationally. For example, the [Fazekas school](https://en.wikipedia.org/wiki/Fazekas_Mih%C3%A1ly_Gimn%C3%A1zium_(Budapest)) is number 6 on [a list of international secondary schools](https://www.telegraph.co.uk/news/0/international-schools-place-oxford-cambridge-study-abroad/) ranked by Oxford and Cambridge offers. The model was copied across the Eastern Europe.
* **mathematical talent selecction:** in addition to secondary schools there is a network of organisations who identify and nurture exceptional mathematical talent through competitions, summer camps, etc. Most prominent is the [Gondolkodás Öröme Alapítvány](agondolkodasorome.hu) (literally translated Joy of Thinking Foundation) which carries on the work of [Lajos Pósa](https://en.wikipedia.org/wiki/Lajos_P%C3%B3sa_(mathematician)#Mathematics_education), an outstanding maths educator who was best known for identifying talent early. These Pósa-camps create a strong network and pipeline of mathematically talented youth.
* **incredible olympiad track record:** The result of the above is that the country punches way above its weight when it comes to mathematical talent. The table below compares the country's success in various international olympiads compared to richer European countries with substantially larger student populations and twice the education spend per student.

| Country | Per Student [Primary+Secondary Education Spend](https://www.statista.com/statistics/238733/expenditure-on-education-by-country/) | [Student numbers]() | IMO [rank](https://www.imo-official.org/results_country.aspx?column=awards&order=desc) | EGMO [rank](https://en.wikipedia.org/wiki/European_Girls%27_Mathematical_Olympiad#Medal_table) | IPO [rank](https://en.wikipedia.org/wiki/International_Physics_Olympiad#Distribution_of_medals) | IOI [rank](https://stats.ioinformatics.org/countries/?sort=medals_desc) |
| ------- |:--------------------------------------------------------------------------------------------------------------------------------:|:-------------------:| -------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------------------- | ----------------------------------------------------------------------------------------------- |:----------------------------------------------------------------------- |
| Hungary | $12,103 | 0.4M | **5** | **6** | **9** | 20 |
| Germany | $24,022 | 4.8M | 10 | >10 | >10 | **17** |
| UK | $24,441 | 2.3M | 11 | 7 | >10 | 29 |
| France | $21,730 | 3.3M | 19 | >10 | >10 | 36 |
| Italy | $21,909 | 1.8M | 25 | >10 | >10 | 45 |

* **lack of competitive research and higher-ed:** The country lacks competitive STEM research and higher education institutes which would take the exceptionally talented students and provide them opportunities to channel their mathematical problem-solving skills to addressing relevant world-class research problems.
* **mediocre research careers:** Most researchers who work on modern, internationally relevant problems have left the country. Research that is carried out at most local labs is mediocre at best. Students engage with mediocre research and often get stuck in mediocrity. They often end up in low-impact careers simply by virtue of working with the first professor they find interesting.
* **studying abroad vs Hungary:** Over the last decade, the best students routinely got places at Oxford and Cambridge, but a great number of talented students still stay in Hungary where degree courses are free. Several of the best students are first-generation college students (I am also a first generation college student from a working class family) who may not consider studying abroad. More recently, Brexit has made Oxford and Cambridge inaccessible even to IMO medalists (costs quadrupled and students can't apply for student loans). Although scholarship options exist, they are scarce, uncertain and restrictive. It is likely even more of the best students will stay in the country for their undergrad, especially so if a high-quality post-secondary programmes existed.
* **colleges of advanced studies:** Colleges of Advanced Studies ([Szakkollégium](https://hu.wikipedia.org/wiki/Szakkoll%C3%A9gium)) are a rather unique model in Hungarian higher education. These colleges are often maintained by universities, have a selective admissions process and offer membership and often dormitory accommodation to the best students studying certain subjects. An example of such College is the [Bolyai College](https://www.bolyai.elte.hu/dyn/english/) which focusses on natural sciences and mathematics. They run seminar talks by members and invited speakers. There are nearly 40 such colleges in the country.
* **government influence in research and education:** Increasing government interference in state-run research and education institutes makes the country's research sector unattractive for the best researchers who can have careers abroad. To read more see e.g. [this](https://www.economist.com/leaders/2021/05/01/viktor-orban-seizes-control-of-hungarys-universities), [this](https://kafkadesk.org/2019/06/29/viktor-orbans-latest-target-the-hungarian-academy-of-science/) or [this](https://www.nytimes.com/2021/06/28/world/europe/hungary-orban-university.html). Recent changes handed leadership of these institutes to undemocratically governed foundations (Members are appointed for life, and only the current members can choose new members). Several researchers I talk to cite this as the main reason for not returning or moving to the country. There are barely any private or independent research institutes in the country.

## 2.2. Guiding Policy

* **differential technological development:** to shape the future, we have to speed up the development of technologies and ideas that are beneficial, relative to those that are harmful. Our focus therefore is areas of AI research that are considered long term beneficial: a mathematical understanding of AI, AI ethics and AI safety.
* **raise the bar for post-secondary education:** for the best students, we can create a post-secondary education option that is on part with the exceptionally good secondary education, and which rivals top universities abroad.
* **prioritise student outcomes:** The purpose of this institute is to nurture and guide exceptional talent, success should be measured agains this, rather than traditional academic indicators such as citations or impact factors.
* **collaborate, not compete:** Our goals will be realised in collaboration with other institutions, we do not compete for talent who have access to similar or better opportunities elsewhere.
* **scale as widely as possible:** We have to scale beyond the top schools, by supporting students who show great promise despite not having acccess to the best education to get into an olympiad team.
* **maintain independence from government/politics:** To attract and retain top talent, the institute must ensure longevity and independence from government/politics. It has to offer an alternative to increasingly government-controlled research and education institutions.

## 2.3. Actions

* **a college of advanced studies for AI research** The institute would follow the Hungarian *college of advanced studies* model outlined above: a research and education institute that confers no degrees but offers supplemental training for high school students and those already enrolled in degree programs. It would have highly selective admission, we would seek out and offer membership to exceptionally talented students, not limited to, but including STEM olympiad contestants. Our curriculum would introduce Master's level lectures, seminars, research skills training, 1:1 mentoring, career advice, research internships, camps and summer schools from a young age.
* **cost-effective elite undergraduate training:** Narrowing the focus on AI and AI safety allows us to create an UG training offering that competes with the best programs globally. The combination of a 3-year CS or Maths degree with the institute's supplemental training program (including 1 or 2 summer internships) could turn out students with outstanding profiles who can go on to selective Master's and PhD programs, or to start work in AI research right away. Compare the costs to top institutions: three years of CS studies in Cambridge cost $230,000. In Hungary, UG courses are free for EU citizens and this institute could bridge the gap in quality of AI teaching and research training for an annual budget of <$35k per student.
* **lecture modules, seminars, reading groups:** We would design lecture modules that complement and build on the typical offering at local universities. For example, we could run modules on:
* *deep learning:* following the [D2L Book](https://d2l.ai/https://d2l.ai/)
* *foundation models and scaling:* less technical module introducing current large language models (LLMs), large vision models, and the idea of foundation models and scaling laws. We would highlight alignment work done at companies like Anthropic or OpenAI.
* *theory of deep learning:* technical module on building an understanding of how and why deep learning works and predicting failure modes, based on a [similar course](https://www.cl.cam.ac.uk/teaching/2122/R252/) I designed in Cambridge
* *AI risks:* non-technical module that illustrates short-term AI risks through case studies (e.g. [deepfake video](https://www.youtube.com/watch?v=pfsdvbacYac) used during Ukraine war or [dual use of AI in drug discovery](https://www.nature.com/articles/s42256-022-00465-9)), introduces AI x-risk scenarios and motivates alignment work.
* *AI technical alignment:* following the [Technical Alignment curriculum developed by EA Cambridge](https://www.eacambridge.org/agi-safety-fundamentals) but introducing more technical material from current reseaarch.
* *AI governance and policy:* Building on the [AI Governance Curriculum](https://www.eacambridge.org/ai-governance-curriculum) developed by the Stanford Existential Risk Initiative.
* *AI ethics*: covering topics such as fairness, interpretability, privacy-preserving ML.
* *AI research skills:* covering topics such as how to read papers, how to critically evaluate them, how to spot potential alignment problems, how to keep up with literature, how to communicate ideas, etc.
* *python for research:* as most university courses and pogramming contests use C++ and Java, we should expose students to numerical progamming in python.
* **research internships for undergraduates** Most AI research internships are targeted at established PhD students, junior students have fewer options. As our focus is on student outcomes, we will offer research internships to younger students, ideally targetting second year undergraduates who already learned enough mathematics to tackle rigorous problems. We will provide close mentorship throughout these scholarships - how to navigate the AI literature, how to read and write papers, blog posts or open-source code. Students studying abroad prefer spend time in Budapest over the summmer, therefore high-quality research studentships will attract STEM students from top universities. We can expose them to AI research and to integrate them to our broader community as well.
* **visiting fellowships for AI, ethics and safety researchers:** This was suggested by David Krueger as a mechanism to engage international AI safety talent in the training of local students. Visiting Fellows would be given an opportunity to spend 6-12 months in Budapest, accompanied by some of their students.
* **researcher positions with long-term independence and security:** If the institute is to attract and retain senior research talent (i.e. those currently in a tenure-track position in academia or researcher position at, say, DeepMindd) we need to create group leader positions that come with secure, long-term funding. We will be creative in offering a variety of different researcher jobs to suit individual goals.
* **grants for outstanding high-school teachers:** We will make fellowship grants for the most talented high-school teachers in the country, who have a track record of enabling talented students. These grants could include funding for:
* a salary supplement
* a grant allowing the school to hire an additional CS or Math teacher, freeing them up to develop differentiated and advanced curricula for the best students and 1:1 teaching.
* funding for afternoon clubs including kids from other schools
* funding for books or equipment
* in exceptional cases, a more significant grant could allow a secondary school to introduce a special maths/CS specialisation program. There are currently [12 such programs](https://matek.fazekas.hu/index.php?option=com_content&view=article&id=78&Itemid=202) in Hungary.
* **retreats and summer schools:** In addition to supplemental lecture series, we will organise thematic intensive courses. These will be partiularly useful in reaching students who study outside of Budapest, and integrating them into the community. We have trialled the AI retreats model this summer with great success.
* **exchanges and placements:** We would like to fund student exchanges, both as a host, and to fund visits by our students to other host institutions. We could host visitors through the [ELLIS PhD and postdoc program](https://ellis.eu/phd-postdoc).
* **an ELLIS Unit:** If our institute is funded, we will apply to become an [ELLIS Unit](https://ellis.eu/units) (needs an annual budget of EUR1.5M). With the exception of Prague, no Eastern European city has such an institute. This would integrate our activities to the emerging European Laboratory for Learning and Intelligent Systems, modeled after the widely successful [CIFAR AI programme](https://cifar.ca/ai/canada-cifar-ai-chairs/#:~:text=The%20Canada%20CIFAR%20AI%20Chairs,retaining%20our%20existing%20top%20talent.) in Canada. This could open up a sustainable funding model from the European Union in the future.

# 3. Existing Groundwork

*-- 8 minute read*

This proposal grew out of side projects I (Ferenc) pursued since 2020. I invested roughly $20,000 of own funds and >300 hours of my time so far. A quick overview:
* **ML reading group:** community of ~40 members, 30 hours of content
* **12-week research scholarships:** supervision + total of $12,000 given from own funds
* **personalized ML training for high-schoolers:** 10-day visit, hosted in Cambridge
* **AI summer retreat** for 12 overachieving teenagers focussing on large language models ([link](http://airetreat.org)).
* **ML book giveaways** to students
* **small grants to schools:** outreach + $500 donated as book purchases
* **EEML:** 2x organiser of East European ML Summer School

### 3.1. Reading Group

Starting in September 2020, I recruited a group of Hungarian students of various seniority to participate in a bi-weekly [reading group on theoretical machine learning](https://hackmd.io/@mljc/HyYC8DzSw). We discuss seminal papers that illustrate core concepts in modern machine learning. In addition to student-led discussions, we invited leading researchers to talk about their work (Neil Houlsby on Large Vision Models, Nicolas Heess on Reinforcement Learning). The group now has 40 members, with about 50-50 gender balance, 9 olympiad medals among them. We accummulated over 30 hours of content. Example participants:
* **Bogi E** is the youngest participant who joined the reading group at age 17. She is from a rural town in Hungary, and has great achievements despite not going to one of the top secondary schools. She participated in robotics competitions since the age of 12, took part in a summer research internship on medical imaging aged 17. She is now a first year student at GeorgiaTech and continues to regularly dial in to each session to fast-track her training in AI. I applied for a scholarship from the EAFunds Long-Term Future Fund on her behalf.
* **Mina** is one of the first members of the group and our expert on privacy-preserving ML and ML ethics. She studied in one of the top CS programs in Budapest. I asked her to join after I found her name looking through a list of scholarship recipients and student conference participants. Based on her engagement with the journal club, I encouraged her to apply for the MPhil programme in Cambridge, which she did, receiving a DeepMind Scholarship. She has an offer to join Moritz Hardt's new AI Ethics lab in Tübingen called 'Social Foundations of Computing' as one of the first PhD students there.

### 3.2. Earning to Give

In order to generate funding for further proof of concept activities, in 2020 I took up a one year advisory position at Twitter, focussing on AI ethics. I did this with the deliberate goal to give my income away supporting Eastern European talent as I announced in the Tweet below:

<blockquote class="twitter-tweet"><p lang="en" dir="ltr">I&#39;m also planning to donate my income from this position. My aim is to help organisations and individuals who create opportunities for talented students in Eastern Europe to get exposed to and participate in world-class research. Will knock on your door <a href="https://twitter.com/EEMLcommunity?ref_src=twsrc%5Etfw">@EEMLcommunity</a></p>&mdash; Ferenc Huszár (@fhuszar) <a href="https://twitter.com/fhuszar/status/1341045408726986754?ref_src=twsrc%5Etfw">December 21, 2020</a></blockquote> <script async src="https://platform.twitter.com/widgets.js" charset="utf-8"></script>

### 3.3. Research Scholarships

In the summer of 2021, I have decided to try supporting summer research scholarships focussing on younger students who often don't yet have many such opportunities. In 2021 I supported two students:
* **Anna M** joined our group at the beginning of the 2nd year of undergraduate Maths studies in Budapest. She went to the top maths secondary school, got a bronze medal at EGMO, and has a near-perfect GPA in Pure Mathematics. With no formal training on probability, she gave an impressive presentation about Wasserstein GANs. In the summer of 2021 she was offered a software engineering internship with Morgan Stanley in Budapest. I offered her the same salary if she turns it down to participate in a research project instead. This eventually lead to a paper on [understanding generalisation properties of neural networks](https://arxiv.org/abs/2111.11542). The project likely helped her get a place in the Part III of the Maths Tripos in Cambridge, one of the most selective graduate Math programs.
* **Anna K** is also an EGMO medalist (1×gold, 2×silver) who is now in her third year of Maths training in Cambridge. Due to COVID, it was difficult to secure an undergraduate reseach opportunity so offered her the same scholarship terms as to Anna M, and the two of them worked on the same [project](https://arxiv.org/abs/2111.11542). Anna also gave [a talk](https://www.youtube.com/watch?v=aduB1owRVQE) at the Mathematics of Deep Learning programme at the Isaac Newton Institute.

In summer 2022, I have been working with four research interns, all outstanding mathematics students, on the theory of sharpness-aware optimization (SAM) and generalisation.

### 3.4. High School Students

Motivated by the success of working with younger students, I started engaging high-school students. As this [blogpost](https://forum.effectivealtruism.org/posts/HcaB2kJKhxJtS4oGc/some-thoughts-on-ea-outreach-to-high-schoolers) outlines, EA outreach among high schoolers can be particularly fruitful.

In April 2022 I will host two 16yo students in Cambridge. My aim is to teach them python programming, pytorch and machine learning, and generally to raise their ambitions to pursue impactful careers:
* **Bogi N** is a student in a specialist maths-CS high school program in Satu Mare, Romania. She is into competitive programming, a member of the Hungarian team in the first European Girls' Olympiad in Informatics (EGOI). She is very keen to learn about atificial intelligence. I proposed a project for her that involves playing with large language models like GPT-3, visualising the universe of answers it would give to a question.
* **Nóri** is also enrolled in a maths-CS special high school programme in Cluj-Napoca, Romania. She started programming robots at a young age and participates in innovation competitions. Her main project is a small robot that goes around, finds trash on the pavement, and then alerts passers-by. She is very keen to build a better understanding of the off-the-shelf computer vision techniques she used, so we will focus on a tash-related vision use-case with her.

I am also starting a newsletter written in Hungarian, targetted directly at interested high schoolers. This week I wrote about deepfakes (with reference to the recent [Zelensky deepfake](https://www.youtube.com/watch?v=pfsdvbacYac) in the Ukraine crisis) and [dual-use of AI on drug design](https://www.nature.com/articles/s42256-022-00465-9).

### 3.5. AI summer retreat

I organised an [AI-themed retreat](https://airetreat.org)/summer camp for 12 of the most promising teenagers I have worked with from a variety of schools and backgrounds. We focussed on large language models (OpenAI provided a $50 credit to use GPT) and the Stable Diffusion image generator.

The students then chose projects to work on, examples include:
* **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(17) and Eszter(18) 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: A language model-based word association game agent:** Hédi(16) and Júlia(18) 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(17) and Csaba(18) 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.

As you can see from the participant bios, the students were from a diverse range of schools outside the capital, and we had more girls than boys in this particular retreat.

### 3.6. Book donations

Even though more and more books are avilable online for free, I believe in gifting people or schools physical copies as a way to encourage them to learn about a topic. In 2021 I bulk-purchased copies of the [Mathematics of Machine Learning](https://mml-book.github.io/) which I have been giving away to talented students. I planning similar purchases of other technicl books like [Kevin Murphy's new book](https://probml.github.io/pml-book/book1.html), and the [D2L book](https://d2l.ai/).

In terms of non-technical books I was planning to buy copies of [The Elephant in the Brain](https://en.wikipedia.org/wiki/The_Elephant_in_the_Brain), a book that had a big influence on my thinking, to introduce young people to the concept of effective altruism. An example of a student who's receiving these books:

* **Júlia** is an EGOI silver medalist from Satu Mare, strong in maths and competitive programming (see e.g. [codechef profile](https://www.codechef.com/users/kideso)). She wants to specialise in areas of CS where she could use her strong maths and algorithmic problem solving skills. So far the only advice she was given is to work on "3D modelling" which needs a lot of maths. Júlia's case highlights the potential of engaging and empowering an exceptional student early to make more informed choices about impactful career paths. She is currently busy taking final exams, in the meantime I am sending her a series of books to expose her to the mathematics of machine learning, and interesting maths challenges in addressing important long-term problems.

### 3.7. Teacher/school grants

I have also reached out to high schools, especially from rural towns, that are starting to build up a track record in nurturing maths and CS talent. This is usually due to the work of a very motivated and talented teacher who is doing great work. I have been offering my help in maintaining and expanding that work. Examples of this so far:

* **Csaba** is a CS teacher in Satu Mare, a rural, Hungarian speaking town in Romania, close to the border with Ukraine and Hungary. Csaba has done excellent work at this school, raising a sting of talented competitive programmers, EGOI medalists and more. I reached out to Csaba in 2021, to thank him for his work and offer my support. He asked for books he can use for teaching programming which are not easily available in Romania. I purchased the school a collection of books and, in addition, bought them online access to [AoPS](https://artofproblemsolving.com/) which is the number one Maths book IMO participants use. Csaba also set up a 2 hour career advice session with the best CS students from the school - this is how I got to know Júlia and Bogi N mentioned above.
* **Tibor Oláh** built a robotics talent workshop in Hajdúböszörmény, a rural town in the Eastern part of Hungary. Among his students was Bogi E, now a first year CS undergraduate at GeorgiaTech, and a member of my reading group. It's difficult to emphasize how crazy unusual her trajectory is. Tibor also set up a Robotics Olympiad for high school students in Eastern Hungary that extends to schools in neighbouring countries. When I reached out to him, he asked for help extending his robotics olympiad more widely, to the whole country. This is not something I could effetively help with so far, but something our Institute could partner with him on.

### 3.8. EEML: Eastern European Machine Learning Summer School

I served as a co-organiser of the East European Machine Learning Summer School ([EEML](eeml.eu)) since 2021. Among my contriutions was introducing a reading group format to the school, to our knowledge the first time this was done in a summer school. The school gives me an opportunity to get more involved with the local AI/ML talent pool.

In the future, I my vision is to shift the way EEML works in recognition of the region's unique challenges. Most ML summer schools target PhD students, while I think we should make the program more accessible to younger students, especially second year undergraduates, and perhaps extend to offering programs to talented high school students, too.

# 4. Leadership

*-- 3 minute read*

This project would be lead by me, [Ferenc Huszár](https://www.inference.vc/about/). Check out my [blog](https://www.inference.vc/), [google scholar](https://scholar.google.co.uk/citations?user=koQCVT4AAAAJ&hl=en), [linkedin](https://www.linkedin.com/in/fhuszar/?originalSubdomain=uk) or [Twitter](https://twitter.com/fhuszar). Below are some higlhights from my CV and my motivations for pursuing this project.

## 4.1. Track Record

* **inFERENCe research blog:** I (Ferenc) am probably best known for keeping a popular research blog, [inFERENCe.vc](https://www.inference.vc/), since 2015. I have written over 100 posts, and I have reached a global audience of about 10k monthly unique readers. Some posts, such as the series on [causal inference](https://www.inference.vc/untitled/), reached a much wider audience. I regularly receive feedback on the blog's usefulness, and some people credit the blog as their inspiration to enter machine learning research. The blog also allowed me to move past the traditional academic publishing model, I often publish original ideas or original interpretations of methods there. Some of my blogposts inspired papers and PhD theses, or are cited in peer-reviewed research.
* **teaching AI to Twitter leadership:** Probably the coolest and highest-impact thing I've done were the weekly 1:1 phone calls with Twitter CEO Jack Dorsey about AI in 2017-19, subsequently extended to all-company AI broadcasts. In these 15-30 minute sessions I talked about fundamental concepts that made modern AI more powerful than techniques that came before. We covered a range of topics from representation learning to how AlphaGo worked. We turned to covering topics of ethical AI: we discussed privacy-preserving ML, algorithmic fairness, societal impacts of recommender systems. In my view, these sessions were instrumental in building support for our fair ML and algorithmic transparency initiatives.
* **Twitter's AI ethics and transparency:** Twitter's organised AI ethics efforts started in 2018 when Naz Erkan and I put forward a case for creating a new team dedicated to algorithmic fairness. In 2019 we [teamed up with UC Berkeley researchers](https://blog.twitter.com/en_us/topics/company/2019/ucberkeley-twitter-ml) and created the team now called META (ML Ethics, Transparency and Accountability) which grew to more than 15 engineers. I spearheaded the team's most important algorithmic transparency project to date on the [amplification of political Tweets](https://www.pnas.org/doi/10.1073/pnas.2025334119).
* **Ukraine Maths & Science Achievment Fund:** In March 2022, in reaction to the invasion of Ukraine, I [started an initiative](https://twitter.com/fhuszar/status/1506301179789287430?s=20&t=On6fleRYZm1k1brybcpXuA) to support outstanding high-school talent from Ukraine. My previous experience working with talented kids in Hungary has helped me design this program, and by July 2022 I reaised $3M from Ken Griffin to create the [Ukraine Math and Science Achievement Fund](https://ukraineachievementfund.org/), which will provide tuituin support the most talented Ukrainian STEM students in the coming years.
* **venture capital experience:** Rather unusually for someone with my research background, I also worked as data scientist at one of the leading VC firms in Europe, Balderton Capital. I had a chance to join partner meetings where partners discussed the dealflow, and where startups came to gave their final pitch for funding. This was an amazing platform for me to learn about how ideas are scaled, how successful teams are built, and to tell apart promising ideas and entrepreneurs from bad ones.
* **entrepeneurial success:** I was Principal Research Scientist of Magic Pony Technology, an AI startup that focussed on deep learning-based video compression technology. The company of 12 researchers and engineers was [acquired by Twitter for $150m](https://techcrunch.com/2016/06/20/twitter-is-buying-magic-pony-technology-which-uses-neural-networks-to-improve-images/) in 2016. Our key differentiator was our product-driven and pragmatic research culture, alumni of Magic Pony went on to take up influential roles within Twitter.
* **curriculum development in Cambridge:** In Cambridge I co-developedd the first-ever dedicated [deep learning module](https://mlatcl.github.io/deepnn/), the [theory of deep learning module](https://www.cl.cam.ac.uk/teaching/2122/R252/) for advanced students, and a short module on causal machine learning.

## 4.2. Personal Motivation

Sidelining a faculty job at Cambridge and moving to Hungary is a contrarian move. It comes with several sacrifices for me (Ferenc) and my family. So why do this?

* **if I don't do this, who will?** A unique combination of factors allows me to do this: my relatively fortunate financial situation which allows risk-taking; my AI research credentials; my knowledge of schools and talent in the country; my entrepreneurial experience in startups and venture capital; and my teaching skills. I strongly believe Hungary needs a high-profile research lab like this to exist. If I don't start this, who else would?
* **leverage:** I believe that in my position, creating and running an institute like this would maximise the impact I can have as an individual. At this stage of my career the collective impact of students I train will eclipse my individual contributions to research and technology. Working with students who have high potential but limited access to opportunities creates singificantly more value than the marginal value of working with students who are already studying at a world-class research university like Cambridge.
* **anti brain-drain activism:** Intellectuals, researchers, educated youth who oppose populist, anti-democratic tendencies in Eastern Europe protest by leaving their country. As a result, there will be even fewer intellectual role models in society who stand up for human values, and fewer institutions that can offer meaningful alternatives for those who want to carry out world-class, internationally relevant work. I believe it is important to create a focal point for those researchers who would like to stay in or return to the country and contribute to building a better society.

# 5. Funding and Budget

*-- 3 minute read*

### Mainline budget

Below is an illustrative spend profile for the mainline ask of $3M for the first five years:


Rough unit economics:
* pipeline "graduates" 5-10 students per year
* each student supported for 4 years
* average annual funding per student: <$20k
* average total funding spent per student: $63,000

*Comparison: cost of a 4-year merit-based undergraduate CS studentship in Cambridge: $310,000 + admin overhead.*

### More ambitious budget

I have also created an illustrative budget for a more significant institute with $7.5M funding (think of this as the potential long term evolution of the institute):


Rough unit economics:
* pipeline "graduates" 50 students per year
* each student supported for 4 years
* average annual funding per student: <$10k
* average total funding required per student: $34,000

# 7. Diversity and Widening Participation in AI

*-- 1 minute read*

* **viewpoint diversity:** The current AI, AI safety and AI ethics communities approach risks from a predominantly English-speaking Western viewpoint. Bringing in people from Eastern Europe (countries with severely limited media freedom, first-hand experience with rising populism, authoritarianism and propaganda, where half the population don't speak English) could provide important viewpoint diversity in perspectives.
* **gender diversity:** The community we have managed to assemble so far happens to have a 50-50 gender balance, in fact there are regularly more women than men in our reading group sessions. This is, in part, thanks to the region doing slightly better in girls' Maths education ([Lippman and Senik, 2018](https://ftp.iza.org/dp11532.pdf)), but it is also due to our successful directed efforts. Our current network is going to be a good starting point for creating a diverse pipeline of talent.