.@0xK4d1r, @arma185, a real-life HFT anon, and I spent the last year building one of the dominant CEX/DEX arb bots on @avax. We share a bunch of the lessons we learned along the way in our post below. Enjoy :-) https://t.co/LyLOFXTdP8

I just posted the 48th edition of my newsletter: * Kyutai decodes spoken LLMs * Irrational markets I: Cartesia's $100M * Dr. ChatGPT retires * Irrational markets II: $25B for a press release * Minimax and NTP is all you need * Booting OpenAI off WA https://t.co/SCWkuMihNP

.@deanbphillips is looking pretty smart right now. The Democratic party and associated media owe him an apology. He saw and campaigned on an unfortunate truth that it took them another painful year to acknowledge.

.@bylloydlee at @insiderbusiness wrote a nice follow-up piece to my blog post on product management at Meta vs. Google. https://t.co/fpiuYsOBxb My favorite bit is that a Google spokesperson felt compelled to respond.

Since returning to Meta in February, I’ve fielded a small deluge of calls from Meta people considering Google offers, Google people considering Meta offers, and non-aligned candidates considering both. I wrote up thoughts on PM at both companies. https://t.co/ktligXPoWI

2/ GPT-3.5 nailed this, but nearly all the open-source models (Llama-70B, Mixtral, Gemma, etc.) and Gemini failed in the same way as Opus. I'm trying to figure out what is hard or confusing about this prompt? I thought it would be slam dunk for everyone.

1/ I finally subscribed to https://t.co/QPo2Nv7esl to get access to Claude 3 Opus and see what all the fuss is about. It failed on my first simple prompt. "I have a csv file containing a header in row 5 and data starting in row 6. How to open in pandas?"

The fact that we don't welcome stars like @rdesh26 into the United States should be a bipartisan embarrassment.

https://t.co/X3BlrcItxg

Every few weeks, I get connected with someone interested in transitioning to product management. While each person is different, every conversation starts with a similar framework, which I wrote down in case it generalizes to others considering the same career move.

4/ A CS professor for example, would probably have an easier time just joining a tech company and funding research out of her own pocket than dealing with government funding agencies, which is I believe what @stephen_wolfram did with Mathematica.

3/ This is significantly below YC, private foundations, and even high earners cashing a paycheck. I recognize that many others have already pointed out how broken the system is for funding academic research, but seeing the numbers broken down blew my mind.

2/ Acceptance rate is 20%. Typically universities take 50% of those funds as overhead. Multiplying these numbers together, it means that the typical professor is only able to raise $200/hour to support her research.

1/ I was just looking up numbers on Federal research grant efficiency for a blog post. The NIH is the largest provider of funding in the US ($45B/year). Its R01 grant is the largest offered (~$600k/5 years). R01s typically run 100 highly researched pages (2 months ?= 300 hours).

I was excited to experiment with Gemini's 1M context window today. I prompted to read Moby Dick and describe all the female characters (needle-in-a-haystack as there are only a few). Gemini refused due to "unsafe content". Bummer that a great product has been rendered unusable.

Thanks to @bmprii for sharing his story about the development of sotatercept, the first new drug developed to treat pulmonary arterial hypertension (PAH) in many years, which will bring relief to patients all over the world. Owen would be proud ❤️.

I spent the last year working on Gemini inside of Google Cloud but never totally understood the business reasons for the substantial investment (~30% of headcount) in renting computers to others when Google is so good at using them for their own means. https://t.co/oa2XBf0sHC

Someone is subjecting @coinbase to a relatively sophisticated phishing attack. 1/ text random numbers that their accounts are compromised 2/ if they respond, call and ask for personal information @CoinbaseSupport, I would try to get ahead of this before your users lose money

I always felt like HumanEval was a good proxy for general LLM performance. For fun, I measured the correlation between publicly reported HumanEval and chatbot arena scores (real human eval): 0.8. Does this mean that one simple metric has 80% of the power of complex human evals?

When I think of existential risks to large parts of humanity: * The next pandemic * Climate change→massive depopulation * Another asteroid AI will be a key part of our solution. So if you want humanity to survive & thrive the next 1000 years, lets make AI go faster, not slower.