A personal update, stepping back but not out After 8 years managing Cosimo Capital, I’m stepping back from day to day portfolio management but still a partner and still excited for Cosimo’s next chapter Navigating crypto winters, face melting bull markets, black swans like Bitcoin dropping 50% in a day during COVID (Luke called me PUMPED UP and said THIS IS OUR MOMENT and it was) taught me more about life, markets, and myself than I ever could have imagined At times getting a crypto fund off the ground was a grind but I lived for it. All the all nighters debugging code or mining rigs, protocol launches in the middle of the night, or reading 4chan/twitter/Discord threads until your eyes bleed... A big part of our edge has been love of the game Some other things I'm proud of - Early DeFi/gaming thesis that crushed - Predicted Tezos supply shock before XTZ did multiples - NFT fund (Pudgies Miladys) great timing - Top miner on multiple Bittensor subnets I’d like to share more thoughts soon. Still investing, advising, looking for new ideas, and enjoying the summer DMs are open. Hit me up if you want to chat

mullet money tip of day: We are firmly in a bear market. No one can dispute this. What is challenging about a bear market is that most market participants feel the need to do something during this time. Cycles play out the same, but the narratives often change. What shouldn't change is your approach as an investor. 1. Find a stable source of income to ride out bear market 2. Stable up and try to earn 4-5% of safe yield while you wait for entries. 3. Buy majors 40-50% off highs. Scale in over a longer term time frame (weeks - months) 4. pay attention to trends. Bear markets are where the winners are built. What am I missing here?

for all the innovation in crypto we never figured out a way gigasend our bags into the S&P500 and that's on us

ai, crypto, etc are a continuation of the internet revolution the fact that the birth of the internet revolution is still in living memory influences our relationship to them

my friends and family ask me "whatever happened to Bitcoin" like its an old friend of mine that moved away

my most useful openclaw agent by far is my fitness tracker basic loop: -i send rambling voice memos on Telegram after a session -it logs all my lifts to a local text file -glazes me a bit -next workout reminds me where i'm at -can recap gains/trends any base model works

ive de-smartphoned several times over the years the things that are painful to give up are Maps and Uber everything else is nice to have except now you often need a smartphone for event tickets, menus, etc its hardly even practical anymore

hilarious LLMs score high on tests without seeing the questions there are shallow shortcuts for this (eliminate outlier answers, etc) but they used advanced reasoning too example of how AI will break every human proxy system involving trust, measurement, compliance etc

fun fact this is the original version

end of an era

been trying to do my AI coding through local, open source models VERDICT: close, but they are just not quite there. not yet suitable for a professional workflow they (Qwen 27B, Gemma 4 31B, etc) too often start by making bad assumptions or wrong tool calls at step 1 and then just go off the rails from there. fail too hard too fast such that course correction is impossible whereas with Claude/ChatGPT the errors are much further downstream and much less frequent. more capable of course correction. etc one trick that does work is having Claude/ChatGPT delegate tasks to local models. have the smarter model spell out the tasks simply, check the correctness criteria (unit tests, etc), and just throw away any work that is bad. try again with updated instructions if you want. given the local models are not paying per token this can actually save a lot of tokens (often at the expense of time) this exercise has given me a better intuition around what kinds of task these local models CAN handle so that i sometimes delegate tasks to them just to save some tokens you could of course operate this scheme using smaller/cheaper Anthropic/OpenAI models, but the Holy Grail is a local LLM (or multiple) that i control, maybe fine-tune, that is essentially free to run and privacy preserving

you don;t hear the word "cryptocurrency" anymore jarring to hear it on old podcasts and such