Devansh Mehta

This account is verified. Learn more

@devanshmehta

prediction markets enjoyer @seer_pm

Pluriverse Joined June 2014

4.5K Following 8.1K Followers 12K Posts

Latest posts 5

Pinned post

Devansh Mehta @devanshmehta Oct 13

3 month update on leaving the EF: i've been working full time with @seer_pm since august!

its been a seamless transition to move from being a client of Seer to now joining their team full time. I originally worked with them on the deep funding pilot, during which we battle tested their infrastructure and smart contracts

somewhere along that journey i got heavily pilled on info finance as a category, so joining them was a natural next step.

the simplest explanation is a two-sided marketplace with traders on one side and information seekers on the other. Seekers pay to create markets on questions they care about, while traders are paid for being right

Vitalik wrote earlier about why AIs are a big unlock for info finance: LLMs can profitably trade tiny, low-liquidity micromarkets that aren’t worth a human’s time.

Another unforeseen reason for their timeliness is how AI is making it cheaper to produce an output than to evaluate. the bottleneck is deciding what deserves our attention, which is where info finance thrives

A few example use cases we’re excited about;

1. Hiring / grant selection

You put out a job posting and get 10k+ applicants. You open a market on whether each candidate gets shortlisted for interview (or, more ambitiously, on the performance score they’d earn in their first 3 months if hired). Your team then reviews applications ranked by market odds.

2. Token listings / partnerships

You’re an exchange with thousands of token listing requests. You run conditional markets on how much trading volume each token would generate if listed, and review from the top down.

3. Market discovery / recommendation algorithms

You’re a trader with an edge in certain kinds of Polymarket markets, but finding them eats most of your time. So you fund a market on whether your wallet will trade polymarket markets. Traders who learn your taste get paid, and you get a ranked list of where to look.

At Seer, I’m focused on growing our trader community, working with information seekers who want markets created, and protocol economics.

DMs are open if you’re drowning in more than you can review or you want to trade these markets!

Devansh Mehta Verified account @devanshmehta Jun 23

Super happy to announce this is my last month at the EF. It's been my home for the last 1.5 years, the revolving door b/w EF and ecosystem is healthy

It's been a whirlwind of a year, here are the things I'm most proud of having achieved in my tenure

1. Collecting a dataset of 3000 rows from protocol experts comparing repos and dependencies for their importance to Ethereum

2. Running one of the first multiscalar prediction markets with @seer_pm that got resolved according to the data collected above, with over $50k in trading volume

3. Hosting 5 data science competitions @JoinPond where we gave over $100k to AI data scientists building models predicting repo importance & sybilness of wallets

4. Writing up a whole academic paper on deep funding that we are submitting for publication at peer reviewed journals and also presenting this week at the United Nations for open source week

5. Helping @clesaege with the controversial validator redirected revenue proposal that got the community abuzz this week

What's next? After a break, most likely going much deeper into the prediction market world. I find it tragic that we can only use offchain PMs and im determined to try correcting that

Onwards and upward, esp grateful to my team mates at funding coordination @vinayvasanji , @thomphreys , @MartinBreiten , @sonkiski & raul who I'll dearly miss and to @VitalikButerin for giving me the opportunity in the first place
View quoted post on X

Devansh Mehta @devanshmehta 17h

Post Devcon observations

-highlight was the number of newcomers to crypto at the event, especially younger crowd as we have a bit of an age problem

- merch hunters were irritating but surely its better than same crypto crowd attending same conferences?

- many i spoke to were planning on attending Lisbon too, so I'd be curious to track number of first time attendees at Devcon india that continue onwards to Lisbon as a success metric for this one

- the venue was the best. Have you ever seen zero lines for the coffee? Food was uniformly excellent and didn't run out either, only issue were less sitting areas to eat the food

- traffic was horrible esp at the venue close time of 7 pm, so it was only possible to attend events near to main venue. On the positive side this may be the first Devcon where it felt like we didn't break the city but were just a small drop in the ocean of mumbai

- wifi was absolutely terrible, not just in the venue but all around. Some team mates didn't come to the venue as they need internet for social media video uploads.

As an Indian i found the bad internet unexpected as we are generally proud of our amazing cheap mobile data, but i guess that's only with local sim cards which most attendees couldn't get?

- the music stage was a vibe

- community hubs were unique to Devcon & something more crypto events should adopt. Ut helped a lot as people in a niche find each other. However not all hubs were equal, so better learnings shared among the hub operators would improve it for next time

- can't speak too much on general programming as i spent all my time at the prediction markets hub. From what i heard, the focus on formal verification was great as young developers got interested in the area

- i think Lisbon is a great next location because there's a strong crypto community which is however rather fragmented. An event like Devcon will help bring them all together

Devansh Mehta @devanshmehta 2d

The security roundtable this morning uncovered the following vulnerabilities to guard against;

1. Security config: Whether to keep liquidity all in one address (like most platforms do) or create one person market. The tradeoff is capital efficiency vs security (vulnerability only drains liquidity in that market)

2. Infinite loop of mint and merge: Migrating from a market on v1 of a contract to v2 or improper rounding of txns can let users do infinite mint & merge bugs that can drain all liquidity

For eg trump-harris is 66.66 and 33.33. you can't give a perfect fraction to users, so the dust should always go to the protocol and not the user, as they could then keep minting complete set to sell forb he dustband run it in a loop to drain liquidity. Ditto for a market on v1 having a different market price than v2 of same market, keep minting both sets of outcomes on the one with lower odds and sell into the version with higher odds

3. Oracle dangers: wrongly configured oracles or even a sudden resolutions of tue market can result in a situation where traders drain all liquidity. For eg, a market on when GTA 6 will release could be a wild west duel scenario where 1st person to read website announcing can simply place drain the liquidity. This problem is worse for an AMM compared to a CLOB, where market orders can be cancelled

Two solutions discussed were an AI bot constantly monitoring internet that pulls liquidity if probability of it already being resolved is over 5%, and a more radical one where the constant in x*y=k constantly changes according to different odds

The next part of the discussion centred around AI use in auditing. The simplest method is simply /goal find all vulnerabilities, a more nuanced one is giving it a list of targets to break but that can trigger safety warnings from frontier modules

The other aspect was combining sherlock for crowdsourcing bug reports with immunefi simulating the bug to categorise.

Someone also mentioned that EF recently required a deposit of around $5 in ETH to submit a bug report for upcoming hard fork just to prevent spam. We discussed a more sophisticated version where a minimal amount of $5 gets you an AI review but a higher bond ensured a human review. Deposits serve as prediction markets of sorts where if the bug is valid it resolves to a higher value while if it's invalid it becomes zero, and if in the grey than market is cancelled & deposit is returned

Devansh Mehta Verified account @devanshmehta 3d

The afternoon session looked at liquidity engines in prediction markets, broadly tackling 2 key questions;

1. Do market makers earn money providing liquidity in prediction markets? Or do they always need an outside subsidy from platforms to remain profitable?

2. How elastic should price of outcomes be wrt trading volume?

The discussion started by differentiating b/w CLOBs and AMMs. We had market makers at the roundtable, who said that CLOBs allow them to be endogenously profitable if there is volume; only if it's a long tail, thin market do they require subsidies from the platform to remain in the green

AMMs on the other hand provide stellar user experience but are terrible for market makers, who uniformly lose money on it. So while AMMs are okay for information finance where a seeker expects to lose liquidity money in exchange for signal, it is less suitable for regular prediction markets

Another part then looked at whether CLOBs can be onchain. A major limitation to doing so is simply frontrunning, market makers need priority treatment in cancellation of orders (which is why it's worse for users than AMMs). Hyperliquids onchain CLOB addresses it by making sure in any block all cancel orders get filled first before any other transactions are processed

We overall concluded discussion on the 1st question on fairly negative grounds, that prediction markets as a web2 product with an offchain clob (like kalshi) are likely the most performant design. Web3 helps with security guarantees by preventing rug pull, esp if the market being traded on is greater than mcap of the platform itself, but otherwise little compels prediction markets to be onchain. One dissenting voice felt the collaboration benefits between market makers would make onchain CLOBs outcompete their offchain variants

The second question was more centred around level of liquidity and how it affects price. The main issue is that too much liquidity results in low price impact while high liquidity makes it go 0-100 very quickly. Appropriately defining the relation between trading volume and price increases of outcomes is a hard math question, esp as it pertains to info fi where price is the product

CLOBs are again a more elegant design for market makers since market makers simply state the price and amount they are willing to take an order against it. It's a bit similar to unusual v3 concentrated liquidity except it's pinpointed at a specific price rather than over a range. Price movements are simply all orders being filled at that point & then having to move up the order book to complete the buy

Overall the most fun part of the discussion was we would state a hypothesis about how some idea would affect market makers, and the suddenly realize we had one in our midst. The biggest benefit of these round tables are simply bringing together all the different stakeholders to advance the state of the art
View quoted post on X

Devansh Mehta @devanshmehta 3d

The afternoon session looked at liquidity engines in prediction markets, broadly tackling 2 key questions;

1. Do market makers earn money providing liquidity in prediction markets? Or do they always need an outside subsidy from platforms to remain profitable?

2. How elastic should price of outcomes be wrt trading volume?

The discussion started by differentiating b/w CLOBs and AMMs. We had market makers at the roundtable, who said that CLOBs allow them to be endogenously profitable if there is volume; only if it's a long tail, thin market do they require subsidies from the platform to remain in the green

AMMs on the other hand provide stellar user experience but are terrible for market makers, who uniformly lose money on it. So while AMMs are okay for information finance where a seeker expects to lose liquidity money in exchange for signal, it is less suitable for regular prediction markets

Another part then looked at whether CLOBs can be onchain. A major limitation to doing so is simply frontrunning, market makers need priority treatment in cancellation of orders (which is why it's worse for users than AMMs). Hyperliquids onchain CLOB addresses it by making sure in any block all cancel orders get filled first before any other transactions are processed

We overall concluded discussion on the 1st question on fairly negative grounds, that prediction markets as a web2 product with an offchain clob (like kalshi) are likely the most performant design. Web3 helps with security guarantees by preventing rug pull, esp if the market being traded on is greater than mcap of the platform itself, but otherwise little compels prediction markets to be onchain. One dissenting voice felt the collaboration benefits between market makers would make onchain CLOBs outcompete their offchain variants

The second question was more centred around level of liquidity and how it affects price. The main issue is that too much liquidity results in low price impact while high liquidity makes it go 0-100 very quickly. Appropriately defining the relation between trading volume and price increases of outcomes is a hard math question, esp as it pertains to info fi where price is the product

CLOBs are again a more elegant design for market makers since market makers simply state the price and amount they are willing to take an order against it. It's a bit similar to unusual v3 concentrated liquidity except it's pinpointed at a specific price rather than over a range. Price movements are simply all orders being filled at that point & then having to move up the order book to complete the buy

Overall the most fun part of the discussion was we would state a hypothesis about how some idea would affect market makers, and the suddenly realize we had one in our midst. The biggest benefit of these round tables are simply bringing together all the different stakeholders to advance the state of the art

Devansh Mehta Verified account @devanshmehta 4d

The one today morning was around new use cases for prediction markets & where privacy fits into that question

One theses that emerged from yesterday was that trading might be one of the few jobs safe from AI. The whole "claude go make me money" meme is probably not going to happen

If that's the case, we can expect more people to start moving into trading. And consequently, new types of markets to start getting created, for eg around who a company will hire or value of an OS repo or whether 2 people end up together if they go on a date

So when there is less need for experts due to AI, what do those displaced experts do? One theses was that they become traders, so recruiters might reskill into a trading mindset to participate in markets on which resume would reach the interview stage or matchmakers into which couples would get together

Which then runs right into the privacy point, which can be of 2 forms;

1. A market like "which territory of Ukraine will russia attack" needs to keep its prices private, since if the Russians know that ukraine knows they would not attack it

@clesaege spitballed an entire design for how to build this out, where traders simply say "buy upto 60%" and are left in the dark if their order has been filled, and at resolution all trades become public

2. A market like "is the reported smart contract bug accurate" needs to itself be private. Since we don't want N Korea to test it out themselves and say yup, the bug was accurate.

In such cases the outcome itself (reported bug) needs to be private. This can be done by either encrypting outcomes & giving decryption keys to whitelisted addresses like company employees to trade, or a sandbox LLM where only agents can access the info and trade but they can't reveal it outside the sandbox

The main people trading on weather markets are financial traders, not meteorologists. Which is sadly a bearish case for information finance. a risk mindset where you are okay to lose money cannot be easily taught to subject matter experts, but traders with a risk mindset can still acquire expertise in new forms of market trading
View quoted post on X

Devansh Mehta @devanshmehta 4d

The one today morning was around new use cases for prediction markets & where privacy fits into that question

One theses that emerged from yesterday was that trading might be one of the few jobs safe from AI. The whole "claude go make me money" meme is probably not going to happen

If that's the case, we can expect more people to start moving into trading. And consequently, new types of markets to start getting created, for eg around who a company will hire or value of an OS repo or whether 2 people end up together if they go on a date

So when there is less need for experts due to AI, what do those displaced experts do? One theses was that they become traders, so recruiters might reskill into a trading mindset to participate in markets on which resume would reach the interview stage or matchmakers into which couples would get together

Which then runs right into the privacy point, which can be of 2 forms;

1. A market like "which territory of Ukraine will russia attack" needs to keep its prices private, since if the Russians know that ukraine knows they would not attack it

@clesaege spitballed an entire design for how to build this out, where traders simply say "buy upto 60%" and are left in the dark if their order has been filled, and at resolution all trades become public

2. A market like "is the reported smart contract bug accurate" needs to itself be private. Since we don't want N Korea to test it out themselves and say yup, the bug was accurate.

In such cases the outcome itself (reported bug) needs to be private. This can be done by either encrypting outcomes & giving decryption keys to whitelisted addresses like company employees to trade, or a sandbox LLM where only agents can access the info and trade but they can't reveal it outside the sandbox

The main people trading on weather markets are financial traders, not meteorologists. Which is sadly a bearish case for information finance. a risk mindset where you are okay to lose money cannot be easily taught to subject matter experts, but traders with a risk mindset can still acquire expertise in new forms of market trading

Devansh Mehta Verified account @devanshmehta 4d

The 3rd roundtable yesterday was around trading strategies in prediction markets. The attendees, many of whom were traders, shared 5 strategies to make money

1. Nothing ever happens : find markets panicking over hantavirus or some other things and bet against it

2. Exploit long tail markets & trust the polls. If you see a market on nigel farage vs count binface at 90-10, buy farage

Appropriately Sizing trades in the above 2 categories is important to not get rekt and also a process for finding them, such as the review channel in polymarket

3. Copy trading: On kalshi you can see track record of bettors, so follow their trades or inverse it if they are bad

4. Oracle manipulation: post a malicious answer to UMA for $750, see bots auto trade based on it, take opposite position. If you profit more than $750 it worked

5. Arbitrage: see when the odds across all platforms for the same prediction markets aren't summing up to 1 and exploit the difference. Esp useful to do it to startup platforms without as much usage as incumbents polymarket or kalshi

The last part looked at AI trading. Most were skeptical that an LLM could ever be a profitable trader by itself ("claude, make me money") so maybe job security for traders is high? One person disagreed as trading is a verifiable task and soon to go the way of the software developer

We did find the following 3 areas where LLM trading without humans could be profitable;

- greenfield markets without existing traders, for eg the deep funding ones on value of a repo. Existing prediction markets mostly appealed to a user base that was already betting on sports or politics and made it a better UX

- use AI to book profits, as otherwise emotions prevent us from selling

- the last trading strategy we came up with in theory was an LLM ascribing a probability that it's an insider trading the market and copy trading if it's above a threshold

We also spoke briefly on advantages institutional market makers have over takers or retail, which concretely is a 150ms advantage. The larger alpha for them is access to datasets allowing them to effectively backtest strategies, as public datasets on prediction markets are not very good esp for high frequency traders or market making
View quoted post on X