Phil

Bingo! Here are the key points we have been making to investors: The future is multi-model and multi-modal. 1. A token is only worth what the task it serves is worth. Falling prices therefore don't mean AI is being commoditized, because most jobs never needed the most expensive model in the first place. This is why Anthropic TELLS users to use Opus/Fable for planning and Sonnet for implementation. 2. Volume and revenue split 80/20 in opposite directions over time. Open-source and low-cost models will handle about 80% of token volume and earn about 20% of the revenue. Frontier models will handle 10–20% of the volume and earn 80–90% of the revenue. That is why the race for the frontier still matters. 3. Cheap models make the structuring work affordable. Embeddings, tagging, and classification used to need a premium model. Open-source models now do that work well enough. So an entire corpus can be processed once and stored in a deterministic, auditable, structured form. 4. Structured data ends the brute-force era. Once the data is organized, there is no reason to reload the same documents into a million-token context window for every query. That approach was always very inefficient. 5. The frontier model does the hard part. Orchestration, tool selection, and multi-step reasoning sit on top of the structured layer. Accuracy compounds across steps: over ten steps, 95% per-step accuracy finishes about 60% of tasks, while 90% finishes about 35%. That gap is what supports premium pricing. 6. Falling blended prices are natural and bullish. A lower cost per token means more data gets processed and more advanced workflows become viable, which raises demand for high-end intelligence. Value accrues to whoever owns the harness, the context, and the data infrastructure underneath. (Only thing I'd caveat is that data from Openrouter and Vercel very much skew toward SMB and solo devs, which are going to use a higher mix of open source vs enterprises.)

16K

Phil @liphengag

2 days ago

@HedgeyeComm So labs are forced to pay billions into perpetuity into training to remain on the frontier and keep premium pricing?

166

liphengag retweeted

0xFaust

@0xFaust12

3 days ago

I think the small, concentrated, high-conviction fund thesis is increasingly hard to square with what has actually happened in the industry. LP dollars have been moving toward passive, systematic, market-neutral and multi-manager structures while traditional LO/Tiger/mid-net L/S has steadily lost share. More importantly, the success of the pod model has already shown that fundamental investing is not some intractible problem. You can break the process into forecasts, exposures, risk budgets, sizing rules while reducing many of the behavioral and portfolio construction problems inherent to the traditional Tiger model. Acadian's work (https://t.co/x7TsUdNlf1) supports these same findings: concentrated portfolios have not produced consistent alpha, while more systematic processes can scale through greater breadth. If quarterly fundamental alpha can be industrialized this way, I don't see why longer-horizon fundamental alpha should remain permanently exempt, particularly when prop firms are spending billions on compute, data, domain-specific transformer architectures while continuing to push their forecast horizons outward. The chart below makes the concentrated portfolio argument even harder because the market has become materially more factor driven through most of recent history. A 10-20 name portfolio can therefore very easily amount to a few large latent factor bets with idiosyncratic noise layered on top. This is also where I think the idea that the answer is a smaller fund with more conviction starts to break down. If the data show that concentrated fundamental portfolios are not producing durable excess returns in an increasingly factor-driven environment, then the natural advantage shifts toward firms capable of extracting many smaller signals, controlling their common exposures and combining them systematically rather than relying on a PM's ability to identify 15 exceptional stocks. I also strongly disagree that fixing this is as trivial as onboarding Arcana and building an automated factor hedging layer. The software interface may be easy; building a risk system that is actually useful in production is not. Off-the-shelf models have materially underpredicted realized volatility in 2026, which pushes you into needing custom factors and covariance estimation and the much harder question of whether the factor structure you are using in your off the shelf model is even correct. Axioma itself distinguishes estimation error from specification error and shows how basic choices around estimation window, frequency, weighting, outliers and autocorrelation materially change the resulting exposures. Then you still need to turn those forecasts and risk estimates into positions. Portfolio optimization is a separate technical discipline involving constraints, turnover, estimation error, transaction costs and the interaction between all of them. None of that is trivial, particularly inside concentrated fundamental funds where many PMs remain skeptical of factor models in the first place. The capacity problem compounds this. Passive and pod capital are heavily concentrated in liquid large caps, so the obvious place for a small fundamental fund to look for less competed longer-horizon alpha is further down the capitalization and liquidity spectrum. But a 10-20 stock portfolio deploying meaningful capital into small/mid/micro caps is going to move price against itself both entering and exiting positions. At that point market impact and optimal execution are part of the alpha model rather than implementation details because they determine how much of the forecast survives into realized PnL. MOSEK treats transaction costs, market impact and portfolio constraints as explicit parts of the portfolio optimization problem for exactly this reason. So the concentrated fundamental shop ultimately ends up needing custom risk models, portfolio optimization, impact modeling and execution research anyway. Which are precisely the capabilities where the large prop firms have spent years building an industrial advantage, which is why I think the more likely endpoint is that they continue moving outward in horizon and subsume more of the longer-horizon fundamental alpha rather than leaving it permanently protected for small discretionary funds. Just my take though, would be interested in @TheStalwart @tracyalloway @__paleologo to opine as well

0xFaust12's tweet photo. I think the small, concentrated, high-conviction fund thesis is increasingly hard to square with what has actually happened in the industry. LP dollars have been moving toward passive, systematic, market-neutral and multi-manager structures while traditional LO/Tiger/mid-net L/S has steadily lost share. More importantly, the success of the pod model has already shown that fundamental investing is not some intractible problem. You can break the process into forecasts, exposures, risk budgets, sizing rules while reducing many of the behavioral and portfolio construction problems inherent to the traditional Tiger model. Acadian's work (https://t.co/x7TsUdNlf1) supports these same findings: concentrated portfolios have not produced consistent alpha, while more systematic processes can scale through greater breadth. If quarterly fundamental alpha can be industrialized this way, I don't see why longer-horizon fundamental alpha should remain permanently exempt, particularly when prop firms are spending billions on compute, data, domain-specific transformer architectures while continuing to push their forecast horizons outward.

The chart below makes the concentrated portfolio argument even harder because the market has become materially more factor driven through most of recent history. A 10-20 name portfolio can therefore very easily amount to a few large latent factor bets with idiosyncratic noise layered on top. This is also where I think the idea that the answer is a smaller fund with more conviction starts to break down. If the data show that concentrated fundamental portfolios are not producing durable excess returns in an increasingly factor-driven environment, then the natural advantage shifts toward firms capable of extracting many smaller signals, controlling their common exposures and combining them systematically rather than relying on a PM's ability to identify 15 exceptional stocks.

I also strongly disagree that fixing this is as trivial as onboarding Arcana and building an automated factor hedging layer. The software interface may be easy; building a risk system that is actually useful in production is not. Off-the-shelf models have materially underpredicted realized volatility in 2026, which pushes you into needing custom factors and covariance estimation and the much harder question of whether the factor structure you are using in your off the shelf model is even correct. Axioma itself distinguishes estimation error from specification error and shows how basic choices around estimation window, frequency, weighting, outliers and autocorrelation materially change the resulting exposures.

Then you still need to turn those forecasts and risk estimates into positions. Portfolio optimization is a separate technical discipline involving constraints, turnover, estimation error, transaction costs and the interaction between all of them. None of that is trivial, particularly inside concentrated fundamental funds where many PMs remain skeptical of factor models in the first place.

The capacity problem compounds this. Passive and pod capital are heavily concentrated in liquid large caps, so the obvious place for a small fundamental fund to look for less competed longer-horizon alpha is further down the capitalization and liquidity spectrum. But a 10-20 stock portfolio deploying meaningful capital into small/mid/micro caps is going to move price against itself both entering and exiting positions. At that point market impact and optimal execution are part of the alpha model rather than implementation details because they determine how much of the forecast survives into realized PnL. MOSEK treats transaction costs, market impact and portfolio constraints as explicit parts of the portfolio optimization problem for exactly this reason. So the concentrated fundamental shop ultimately ends up needing custom risk models, portfolio optimization, impact modeling and execution research anyway. Which are precisely the capabilities where the large prop firms have spent years building an industrial advantage, which is why I think the more likely endpoint is that they continue moving outward in horizon and subsume more of the longer-horizon fundamental alpha rather than leaving it permanently protected for small discretionary funds.

Just my take though, would be interested in @TheStalwart @tracyalloway @__paleologo to opine as well

373

660

131K

Phil @liphengag

3 days ago

@LeoPTrudel Bearish for humanity, so... bullish?

liphengag retweeted

Mark Zuckerberg

@finkd

3 days ago

Opening access for developers to build Muse connectors. You bring the API -- Muse brings the agent, the browser, and the context of what the person actually wants. People reach your service just by asking for it, and their agent takes it from there. New connectors are live today. Come build with us. https://t.co/o6oTf1Sj9y