i build ai products, do economics, and ponder the intersection of both
Singapore
Joined September 2011
- Tweets365
- Following764
- Followers582
- Likes1.2K
hey everyone! I made a new fantasy f1 league that you'll find interesting! this is the world's first fantasy game where you can use text to predict literally anything you imagine! check it out before the season starts!
teapaw.org/thamsuppp/status/19001…
🏎️ Ahead of the new #F1 season, I'm launching
🏁 F1 Fantasy: Freeform 🏁 the world’s first text-based fantasy sports prediction game, with AI-powered scoring!
Forget predicting just race results with #f1fantasy - now you can call any scenario you can imagine: first corner crash, unexpected safety car, multiple red flags!
Our LLM scoring system rates your predictions by accuracy vs official race reports. Give it a try with last year's races! (my friends tried to game the AI 👇- can you?)
Can you predict what happens on Sunday's #AustralianGP? Join thousands of fans and earn bragging rights as the ultimate F1 prophet!
hey everyone! made a new fantasy f1 league that you'll find interesting! this is the world's first fantasy game where you can use text to predict literally anything you imagine! check it out before the season starts!
teapaw.org/thamsuppp/status/19001…
🏎️ Ahead of the new #F1 season, I'm launching
🏁 F1 Fantasy: Freeform 🏁 the world’s first text-based fantasy sports prediction game, with AI-powered scoring!
Forget predicting just race results with #f1fantasy - now you can call any scenario you can imagine: first corner crash, unexpected safety car, multiple red flags!
Our LLM scoring system rates your predictions by accuracy vs official race reports. Give it a try with last year's races! (my friends tried to game the AI 👇- can you?)
Can you predict what happens on Sunday's #AustralianGP? Join thousands of fans and earn bragging rights as the ultimate F1 prophet!
🏎️ Ahead of the new #F1 season, I'm launching
🏁 F1 Fantasy: Freeform 🏁 the world’s first text-based fantasy sports prediction game, with AI-powered scoring!
Forget predicting just race results with #f1fantasy - now you can call any scenario you can imagine: first corner crash, unexpected safety car, multiple red flags!
Our LLM scoring system rates your predictions by accuracy vs official race reports. Give it a try with last year's races! (my friends tried to game the AI 👇- can you?)
Can you predict what happens on Sunday's #AustralianGP? Join thousands of fans and earn bragging rights as the ultimate F1 prophet!
Isaac Tham retweeted
I’ve created a dashboard that projects potential GDP losses for countries if ongoing tariff escalations trigger a global trade war.
I’ll be updating it soon with the latest data.
alashkar.pages.iu.edu/Global…
ALT https://alashkar.pages.iu.edu/GlobalTariffWar.html
wow! @Google Gemini 2.0 with Apps summarizes YouTube videos rapidly and amazingly. though it only uses transcripts and metadata, and doesn't have vision capabilities yet
youtube.com/watch?v=yPvoKz6t…
for reference, this is the video i asked Gemini to tell me the number of laps from
wrote about practical evals for LLM summarization recently, thanks @TDataScience for sharing!
Summarization is uniquely challenging to evaluate:
🧠 tough to create synthetic reference summaries, unlike Q&A and RAG
🤔 a "good" summary is subjective and qualitative
📉 many eval methods are either too simple (ROUGE only cares about word overlap) or too complex for practical builders to implement (neural eval models like Seahorse or SummaC are GBs big)
A useful framework is to reframe summarization as a precision & recall problem
🟢 Recall: Does the summary retain important points from the source?
🔴 Precision: Does the summary avoid hallucinating info not found in the source?
This essentially turns summarization evals into a classification problem ✅
@confident_ai DeepEval’s neat Summarization Metric offers a decent implementation of this idea:
Coverage score = recall
Alignment score = precision
I improved on their metric to:
🌟 Use more complex questions to eval coverage
📜 Consider the entire source when judging alignment
⚖️ Balance precision & recall with an F1 score
Lastly, I talk evals for other good summary qualities:
✂️ Conciseness: metrics like entity density, repetitiveness, and sentence vagueness can help
🔗 Coherence: cosine similarity between nth & (n+2)th sentences measures how well the summary flows
Hopefully, this is a practical yet research-backed approach to summary evals for AI builders! 💡Full article in quoted tweet!
🙌 Many ideas for summary evals were inspired by really informative articles by @eugeneyan , @HamelHusain and @ivanleomk 📖👇
Summarization is one of the most practical and convenient tasks enabled by LLMs. @thamsuppp shares an easy-to-implement, research-backed and quantitative framework to evaluate summaries in his new article.
towardsdatascience.com/how-t…
📉 switching to my economist hat – I was profoundly enlightened by Professor Atif Mian’s @AtifRMian talk on the Global Debt Supercycle at the Monetary Authority of Singapore yesterday.
He outlines a dangerous debt trap the global economy is stuck in. Breaking it down:
Phase 1: The Private Debt Cycle ('80s - 2008)
💰 Rising inequality leads to the Glut of the Rich - the rich a much higher savings rate! so overall, savings rise and consumption is depressed
To keep the economy growing, the non-rich to borrow even more, fuelled by ever-lower interest rates
⚠️ This debt-fueled growth continues... until 2008!
Private debt hit a wall - interest rates hit zero lower bound - maximum borrowing capacity reached
Debt was forced to unwind, non-rich households deleveraged, and the economy risked permanent stagnation
Phase 2: Enter the Government (post-2008)
🏛️ Governments stepped in to save the situation.
They could run deficits to stimulate growth while the private sector deleveraged
This worked post-2008 as public debt could grow safely as long as interest rates (r) < growth rates (g)
🚨 BUT if public debt grows too high (as many fear now), we risk fiscal dominance:
Bond vigilantes demand higher rates, and debt spirals out of control
The alternative? Austerity and raising taxes🥴 But fiscal tightening is hard—politicians hate it.
So there’s not much hope in the long-term.
💡 Professor Mian’s takeaway: Monetary policy isn’t neutral! It shapes real economic growth and the “neutral” interest rate.
His solution?
✅ Focus on supply-side investments: fund productive activities that drive sustainable growth.
❌ Avoid short-term, debt-driven credit booms.
🤔 I wanted to ask a question (but was dialing in)
🤖 How does artificial intelligence - and the rapid growth it brings - affect this macroeconomic view?
My take below 👇
AI would solve Prof Mian’s supply-side concerns, and we could see higher growth but lower interest rates!
🤖 AI can unlock massive productivity gains → boosting real GDP growth (10%+ p.a. optimistically)
Higher growth lowers debt-to-GDP, letting governments borrow more without spiraling into crisis
Sounds great, right? Well, not so fast…
In the short term, rapid AI adoption could exacerbate the debt problem
Huge demand for AI investments in the private sector (and soon from govts as AI becomes a security / geopolitical priority)
🚨 This could worsen govt deficit spending and fiscal concerns, before AI’s productivity gains are realized
🌊 Next, powerful AI will displace labour, benefitting capital which owns AI models and compute, further worsening inequality
💰 This means more savings, less consumption (as the income of most people - “losers of AI” - will fall),
Combined with returns to capital eventually saturating, will lead to lower interest rates, and we’re back in the debt trap!
Déjà vu? 😮💨
✨ To avoid this fragile, debt-driven growth model, here’s the better alternative:
💸 Aggressive redistribution funded by compute / capital / wealth taxes to sustain consumption for the broad swathe of “losers of AI”
With AI’s huge growth, there’s plenty of wealth to redistribute—and doing so might be politically necessary to prevent unrest ✊
TL;DR: AI has the potential to supercharge global growth 📈, but it risks exacerbating inequality and economic fragility.
The key? Channel AI’s gains into sustainable, equitable growth - if not for moral, at least for economic reasons!
...before it becomes just another chapter in the debt supercycle 📉➡️📈
Thoughts? @AtifRMian and others, would love your takes! 🧠
Had a thought-provoking discussion at Lorong AI’s Paper Club with @gabrielchua @PradyuPrasad and more, diving deep into Meta’s Large Concept Models (LCM) paper! 🤓📄 Haven’t gone this deep into theory since college
What’s LCM all about? In essence:
✨ LCM is a novel architecture that predicts concepts, or continuous embeddings in a latent concept space, instead of tokens
⏳ It uses SONAR, a sentence-level encoder/decoder embedding model, to convert text → sentence embedding space. LCM then works on this space
Key takeaways:
🌈Unlike transformers (softmax over discrete tokens + token sampling), LCM generates continuous sentence embeddings - hence diffusion models come into play (alternatively, can quantise into discrete)
This clearly illuminated diffusion vs transformer models for me - inspiring me to learn more about diffusion 🌊
🌍🔤Multi-lingual & multi-modal benefits: Though trained on English, LCM performed well 0-shot in other languages - as it operates in a language-agnostic concept space
Of course, this requires a good multi-lingual embedding model of SONAR
Nevertheless, with a common concept space, LCM’s inference is presumably more efficient than multi-lingual LLMs, which need to learn/store concepts in multiple languages.
📚LCM shines at summarization & text expansion, as expected for a model dealing with concepts.
I was personally interested in their summary eval metrics, like: ROUGE-L, coherence (cosine similarity between n & n+2 sentences), repetitive 4-grams, and other neural summarization evaluators
Some described LCM as a “VAE with fancy data prep” 🤔 – having loved VAEs back in college when studying deep learning, I enjoyed the parallels - e.g. round-trip L2 score, reminiscent of reconstruction loss
Lingering questions:
❓ Would LCMs hallucinate more or reason better than token-based models like transformers?
❓ Does the concept embedding space compress/lose information?
❓ how would new architecture ideas like this feed into future AI model development?
Similar work: ConceptBeam - i'll look into that! 📝
All in all, this fascinating discussion got me super excited about deep diving into architectures again! 🌌✨
💡attended a talk "Towards Pro-Human AI" by MIT's Prof Asu Ozdaglar today (was treated as an MIT alum!🙂)
her core idea: AI should expand, not hinder, human flourishing
right now, we're at risk of AI doing the opposite! thru excessive automation (lol she cited her husband), spreading misinformation, and discouraging human experimentation! 😟
the last one was personally fascinating! in short, people rely too much on ChatGPT, instead of engaging in meaningful discourse online or creating solutions themselves.
this results in “informational herding”, stifling diversity and creativity. 🧠
what can we do? focus on complementary AI - tools that augment, not replace, human decision-making.
imo, this calls for a thoughtful analysis of AI vs human strengths, and how to design seamless interfaces to leverage both.
the field of human-computer interaction has been pondering this for decades!
reminds me of how the AI world has evolved toward “canvas” or “copilot” interfaces where AI & humans co-create!
will the future of AI enhance or hinder our flourishing? 🤔 we don’t know yet... but we can carefully design it in the right direction!
launching aiRead - frictionless reading & notetaking for genuine learning!
unlike other AI PDF tools, aiRead helps you enhance, not avoid, your reading - that's when you truly learn!
3 features you’ll love 👇🏼
🖍️highlight as you read - we generate personalized summaries based on your highlights
check out airead! DM me to get code for free beta access
built together with @Jos3ph_s1a
airead.me/
longer article about my inspiration to build this
medium.com/@iztham/airead-yo…
