Eric La Rosa

Person. Prev: BEA, Pivotal, Built AI platform that predicts stock prices & trades algorithmically.

[Eric La Rosa @ColapesceShrugs]

The moment whatever happened

[Eric La Rosa @ColapesceShrugs]

Was randomly tuning into White House Correspondants dinner and about minute ago President Trump and others on stage hit the ground and military types drew weapons on the stage looking through the crowd and the secret service guys came through the crowd. The feed just got cut after that, kind of odd #whitehousecorrespondantsdinner

[Rohan Paul @rohanpaul_ai]

Turns out, it's all in how you ask the question. Prompting is super-powerful.

A single finite Transformer can compute anything computable, just by changing its prompts

Mathematical proof that prompting makes Transformers universal computers

Original Problem 🤔:

The field lacks theoretical understanding of how a single LLM can perform multiple tasks through prompting. While empirical success exists, we don't know the fundamental capabilities and limitations of the prompting paradigm.

-----

Solution in this Paper 🔧:

→ The paper introduces a novel computational model called Two-tape Post-Turing Machine (2-PTM) that can efficiently simulate Turing machines

→ They construct a finite-size decoder-only Transformer that can execute 2-PTMs through chain-of-thought steps

→ The solution uses ReLU activation, layer normalization, and causal attention to implement boolean operations and equality checks

→ The paper develops special encoding schemes to represent arbitrary computations in finite-sized prompts

-----

Key Insights 💡:

→ Prompting is Turing-complete - a single finite Transformer can compute any computable function with the right prompt

→ The model doesn't need to memorize all functions or have answers embedded in prompts

→ The solution achieves nearly same complexity bounds as unbounded Transformers

→ Chain-of-thought steps are necessary for universal computation

-----

Results 📊:

→ Can compute any TIME(t(n)) function within O(t(n) log t(n)) chain-of-thought steps

→ Achieves O(log(n + t(n))) bits of precision for any length-n input

→ Can decide any P language within log-precision

[Eric La Rosa @ColapesceShrugs]

@pmarca "may you live in interesting times"

[Eric La Rosa @ColapesceShrugs]

@realMattKParker hey man i took my freedom and responsibility very seriously, always, and if nothing else i have been authentic to a point of self destructiveness. this is very painful stuff, in truth, if you can't stop looking intently

[Marc Andreessen 🇺🇸 @pmarca]

Enjoy! 💪🇺🇸🔥 https://t.co/UWFfY5nV1o

[Eric La Rosa @ColapesceShrugs]

@pmarca last one you did on JRE was great, looking forward to this one

[Jurrien Timmer @TimmerFidelity]

The Fed seems to have threaded the needle just right (so far), starting its cycle of policy normalization from restrictive (5.25-5.0%) eventually to neutral (3.5-4.0% by my estimates). The rate of inflation continues to edge closer to the Fed’s target, while the labor market has worked off its pandemic-era excess labor demand.
Friday’s employment report, following an earlier JOLTS report, shows the labor market potentially reaching a state of balance. Now the trick is for it to stay there. We know from history that the labor market pendulum tends to swing from one extreme to another, so for it to stop right here at the zero line would be quite an achievement.
Following the strong jobs report last week, the market has walked back its expectations of multiple jumbo rate cuts. We’re back to a more normal trajectory of 25 bps cuts and a higher terminal point (which makes sense). As a result, rates backed up last week, with the dollar index following suit. The dollar continues to be held captive by interest rate differentials.
Nevertheless, with the Fed cutting rates, the PBoC firing a big bazooka, and other central banks around the world also in easing mode (except for the BoJ), it’s no surprise that the global liquidity cycle is on the rise again (and with it stocks, gold, and Bitcoin). The global money supply (as estimated by Bloomberg) is at an all-time high of $107 trillion.

[Marc Andreessen 🇺🇸 @pmarca]

If this actually happens (🤔), I will revise my view of what is possible substantially to the upside.

[Eric La Rosa @ColapesceShrugs]

@realMattKParker yes it is in that scene. when he considers what he might have to do to protect his son, it was just too much for me to even consider someone having to think this

[Jared Friedman @snowmaker]

These are the Honeycomb founders, @azliu0 and @theRealIshank. They're 19 year olds who dropped out of MIT to build this.

[Bravos Research @bravosresearch]

This is the most concentrated market in the past 40+ years

Just the top 5 companies now make up 28% of the S&P 500

Let that sink in

[zerohedge @zerohedge]

US corporate bankruptcies are the highest in 7 years

[Eric La Rosa @ColapesceShrugs]

@elonmusk
I spent three years in my garage building an AI platform for algorithmic trading before I got to the point where tests satisfied me.  I traded a volatile tech stock using only buy/sell signals for a full quarter and saw ~18% profit, despite that stock dropping over 50% and the nasdaq dropping 10% in the same period.

Random forests were used initially due to the amount of data points I had when I started, but I could now increase the data volume and replace models with a Transformer architecture to see how results improve.  If you were to fund me and a friend, we could be used as a test case for using many numerical time series variables to predict another time series variable, with xAI technology at the core, then make specific prescriptions of actions to take (trading) based on those predictions.

Funding came out of my own pocket at the start, as I had just left https://t.co/9DKQ6WvF8N and their IPO was reasonably good.  When I tried to get conversations with Angels and VCs around the bay area I got almost nowhere, they don't seem to fund much for wall street or hedge funds.  I've now spent as much as I can afford to in self-funded mode, on my way to going broke as many dedicated entrepreneurs are capable of doing.

I wrote this up on LinkedIn as a high level overview: https://t.co/rGji5Seo5E.

Please contact me if this sounds interesting.

Thanks,
Eric La Rosa
https://t.co/p3rLuccDDU

[Eric La Rosa @ColapesceShrugs]

@pmarca for me it would be "research more" in general, which might include coding experiments, but that really nails a sense of despair that can hit you at times

[zerohedge @zerohedge]

Atlanta Fed latest Q3 GDP estimate 2.9%, unch from August 6