I suppose hyper-literacy, superintelligence, and existential dread are co-entangled

on the false promise of quantum simulation the power of a quantum compute simulation system given a number of qubits distills to [qubits] = [discrete elements being modeled]^(1/[interaction variables between each element]) The major promising commercial buckets for quantum

tldr meaningful commercial milestones still sit 10+ years out even under optimistic assumptions

Large language repositories Meta is feeding LLaMa with 15 trillion tokens but has a proprietary language dataset that is >10x larger that it is holding in reserve. Reddit is getting $70m per year for its language data; xAI via 𝕏 has access to a repository 25x that size.

Given our emerging understanding of microplastic infiltration We should probably put an immediate stop to all plastic recycling It is an effective subsidy to plastic manufacturing that results in more CO2 emitted per plastic molecule created and likely puts plastic into more

A well-understood bias in science incentive structures—null results don’t merit publication—overcome in this instance Was debating whether AI-assisted science is likely to lead to less scientific agility (conventional wisdom becomes even more entrenched) or more (the cost of

This weekend, the @xAI team brought our Colossus 100k H100 training cluster online. From start to finish, it was done in 122 days. Colossus is the most powerful AI training system in the world. Moreover, it will double in size to 200k (50k H200s) in a few months. Excellent

watched it happen and the rate of progress is still unbelievable

Probably a 4 year build if relying on power/datacenter vendors, contractors, and consultants xAI gets it done in 4 months Given the dramatic performance improvement rate in AI, velocity/urgency is the key determinant of success

Size matters Facebook has ~10x the proprietary language data on database as was used to train the LLaMa models In images they have 20x more than that Instagram and Youtube have 2x more that in uploaded video And yet Tesla's data capture-ability dwarfs all (at 20x more again)

microsoft co-pilot for teams/office: is there some way to use it that makes it actually useful? None of the integrations with word/excel/powerpoint seem to do much of anything (other than waste the time I spend trying to query/task it) Am I just a bad user?

Reason to be skeptical about Apple's ability to deliver a useful consumer-experience with Apple Intelligence Microsoft, with access to OpenAI's models all the way along and massive compute resources, hasn't cracked it. Nor has Google, with TPUs and a massive data abundance.

big data vs BIG data GPT-4 class models are trained on 15 trillion tokens I put the UK biobank's whole genome database of 500,000 individuals at 400 trillion tokens Reasoning follows: 800 million tokens per genome on a similar order of magnitude token vocabulary size to the

GPT 3 class models: AGI is 30ish years away GPT 4 class models: AGI is 10ish years away GPT Next class models: ...

Prior Replit ⠕ thesis: Everyone codes on Replit New Replit thesis: Replit codes for everyone