Get live statistics and analysis of christian's profile on X / Twitter

Technology is good so long as it allows you to go deeper into reality, not flee from it

If 80% of tasks in a job can be delegated, you would actually expect the scope of the role to expand since that same person can take on more responsibility with their surplus time and attention.

More powerful model. Switched basically 100% to Claude when 3 Opus dropped and haven't looked back. Besides that ... the ability to put a custom GPT behind a phone no. and get access to all the raw data.

Can someone explain why the small models are such a big deal? Are there really that many arbitrage opportunities whereby massively horizontally scaling some search/compression process over text yields net margin? I reach for Opus 99% of the time... Examples appreciated!

Long detailed prompts are a waste of time unless they were constructed by a top 1% expert for the particular task (they implicitly know what part of the model to sample, which is best derived from live Q&A, not typing) and contain a whole bunch of gold standard outputs (and, at best, the exact inputs from which they were derived).

Same vibes

For reference, if you wanted to run inference on literally every single word that was spoken or heard in the United States in a 24-hour period, it would cost you less than $200,000.

"Known but not well-indexed" is an interesting way to think about where LLMs are super helpful... Lotta that in knowledge work outside of SWE w complicated rules and regs.

I'm only getting a Vision Pro once it's a proven second monitor alternative. The idea of having a 500-foot canvas I can talk to and 3D objects I can manipulate is tempting but if the weight is too great, I may have to hold out for a V2/3. Ideal workday is 2-4 hours in an immersive, AI-powered digital space, absolutely crushing work at levels of productivity that were previously impossible and then being almost completely unplugged from then on. Once this becomes feasible, we are going to be retrospectively horrified by how inefficient and expensive cubicle-style office workplaces were. Real remote work hasn't even been tried.

The framework is well past the "cargo cult" phase, not unlike uppercase "Agile" or any other management framework from the 90s 'til today... That said, getting a lean, high-performing team to derive the right objective (and measure it three ways) based on proprietary data and a collective exercise in imagination and reason — and then ruthlessly pursue it over the course of 3–4 months, with discipline around measurement, sharing insights, and running risky experiments every week will absolutely yield value. Needs the be the right context (crisis, enabling technology, or energetic/visionary leader), but it does work. (I've been doing this for years and refuse to work with clients of any size who can't commit to a single O and 3x KRs.) The order-out-of-chaos approach is not for everyone, and attempts to implement it with folks who aren't operating on the frontier (the c-suite de facto should be, although that is unfortunately not always the case) is almost certain to be a misfit — hence, the tendency toward degradation into cargo cult rituals when there's an attempt to roll this kind of thing out across more than one team.

When are we going to see AI Labs start really differentiating? OAI -> consumer, accessibility, speed, personalization, mobile compatibility? Anthropic -> power? I find myself reaching for Opus when I need the galaxy brain capabilities, but also loving voice + GPT on iOS...

Something I've noticed that's really interesting about moving to this asynchronous by default working style is that not only are you able to pull outcomes or actions into the present instead of having to wait to have a meeting or schedule it sometime in the future, so your cycle times are basically reduced, but you're also able to find a nice fit between your own natural energy levels and your queue of tasks. There are heuristics that you can use to block time, let's say in the morning, to do deep work and estimate when you might be in the best mindset to do something, but even that's not a hundred percent predictive. So, instead being able to crush the things that need to get done once you've decoupled them from live meetings, when you have the energy, and then attend to those other items later when you're tapped out, has this interesting effect where you're constantly pulling things into the present. And it also makes work significantly more enjoyable so you both get the upside of moving things forward without having to do a live meeting which, for example, you can do with voice agents to collect information asynchronously, but also it avoids the downside of somebody being wiped out or low energy or whatever the case might be when you finally get to that scheduled meeting. Lots more to think about here, but this is a pretty revolutionary way of working and it's boosted my own productivity substantially.

Can someone explain why the small models are such a big deal? Are there really that many arbitrage opportunities whereby massively horizontally scaling some search/compression process over text yields net margin? I reach for Opus 99% of the time... Examples appreciated!

Makes sense, but individuals can access Opus quite a bit for less than a dollar a day! What's the most compelling example you've come across where the small models are a game-changer?

This might be AI's most important feature. Good questions invite focus and ultimately action. So, if these systems are good enough compressors, then with enough context data - with enough high fidelity proprietary data in their context window that's salient to your life and your goals, along with the breadth of its pre-training set and whatever other sources it's able to semi-automatically retrieve from. Then, given some specified higher-order goal, it should be able to prompt you into thinking about things in the right way and then taking the right action. The search engine that asks you questions. The LLM that prompts you.

Is there such a thing as a low-trust, high-alignment organization?

Joining a livestream next week to bust AI myths & share real-world use cases where the tech is actually super valuable... Which one should I demo in real time?

If OAI is serious about search, they should make custom GPTs the primary object that's retrieved, vs. traditional webpages or multimedia. Would be in line with their original "rev share" announcement and a major differentiator vs. Google (or Perplexity for that matter).

We are 2 _______ away from valuable, feasible, and viable AI agents capabilities (eg can produce reliably good results after taking 10+ steps) in a consumer product

If 80% of tasks in a job can be delegated, you would actually expect the scope of the role to expand since that same person can take on more responsibility with their surplus time and attention.

I'm only getting a Vision Pro once it's a proven second monitor alternative. The idea of having a 500-foot canvas I can talk to and 3D objects I can manipulate is tempting but if the weight is too great, I may have to hold out for a V2/3. Ideal workday is 2-4 hours in an immersive, AI-powered digital space, absolutely crushing work at levels of productivity that were previously impossible and then being almost completely unplugged from then on. Once this becomes feasible, we are going to be retrospectively horrified by how inefficient and expensive cubicle-style office workplaces were. Real remote work hasn't even been tried.

True voice-to-voice will be a big deal, maybe bigger than the ChatGPT launch. The fact that I could tap that icon from my lock screen and immediately start speaking with my computer is, in and of itself, incredible. And then it'll only get better over time. First, with integrations and your private data, the quality of the conversations will go up as it gets better at knowing more about you. Then eventually, and I don't know how long this could take, I suspect it'll be years or even a decade, it'll start taking actions on your behalf. And with each release, each step change and backend model capabilities, every single conversation you have with this thing will just get better and better and better. Thinking about the implications for business, the consumer internet, and all the interesting knock-on effects that might happen as a result of this fundamental shift in personal computing are really worth thinking about. Like the mainframe to the PC or the PC to mobile.

When are we going to see AI Labs start really differentiating? OAI -> consumer, accessibility, speed, personalization, mobile compatibility? Anthropic -> power? I find myself reaching for Opus when I need the galaxy brain capabilities, but also loving voice + GPT on iOS...

Solve problems via 1:1 consulting first, get good, inject software into parts of your process, generate a nice inflow of working capital to fund incremental R&D... So many problems to solve, inefficient processes, and people willing to pay if you take the time to understand their business. Don't need a product out the gate IMO.

AI-as-exocortex that extends memory, imagination, and (soon?), reason.