Anthropic's primary Capex partner is AMZN. AMZN is presently willing to drop 200 billion a year into Capex for compute to rent to anthropic and others. This 30 billion only needs to fund their rental rates - unlike openai and google who need to put in the upfront capex for their compute as MSFT stopped footing the bills.

An interesting question is whether anthropic's capex needs may grow to the point that they can take down AMZN with them should they fail.

Does it really? Or does it become a panacea to make people feel like they’ve done something when they haven’t?

Perhaps if these non-solutions didn’t exist to appease our fears then there would be more pressure for real solutions.

I don’t understand why most cloud backend designs seem to strive for maximizing the number of services used.

My biggest gripe with this is async tasks where the app does numerous hijinks to avoid a 10 minute lambda processing timeout. Rather than structure the process to process many independent and small batches, or simply using a modest container to do the job in a single shot - a myriad of intermediate steps are introduced to write data to dynamo/s3/kinesis + sqs/and coordination.

A dynamically provisioned, serverless container with 24 cores and 64 GB of memory can happily process GBs of data transformations.

The history of software has been that once it becomes cheap enough for teams to flood the market with “existing product” + x feature for y users. The market consolidates around a leader who does all features for all customers.

I’d bet that we skip SaaS entirely and go to Anthropic directly. This means the ai has to understand that there are different users with conflicting requirements and that we all need the same exact copy of the burn rate report.

No known mechanism, but cross species checks would imply that the schedule was evolved and has some control mechanism.

Species that evolved before the Devonian period tend not to age and instead grow through their entire lives. There is no mechanistic understanding for the wild variation in species lifespans.

So the natural question in these studies is what would happen if we simply told the muscles not to age this way. It’s plausible that this aging schedule evolved due to other factors independent of the biological constraints. It’s also plausible that evolution removed some other important components for longer lived stem cells.

Interesting, the Devonian also appears to be the period at which fish started sporting limb like appendages and muscle structures, and other animals started to explore land. Perhaps unlimited body growth doesn't work well for animals not entirely supported by water.

Interestingly, the limbless boa constrictor's growth slows, but never stops.

I misread that as the "Denisovan period" and found it interesting that in addition to Homo Floresiensis Hobbits, there might have been arbitrarily large Denisova Hominin giants. Oh well.

I will have to be satisfied with Andre the Giant.

Do you have any advice for running this in a secure way? I’m planning on giving a molt a container on a machine I don’t mind trashing, but we seem to lack tools to R/W real world stuff like email/ Google Drive files without blowing up the world.

Is there a tool/policy/governance mechanism which can provide access to a limited set of drive files/githubs/calendar/email/google cloud projects?

A Google project with capped spend wouldn’t be the worst though, 20 dollars a month to see what it makes seems like money well spent for the laughs.

Not going to lie… reading this for a day makes me want to install the toolchain and give it a sandbox with my emails etc.

This seems like a fun experiment in what an autonomous personal assistant will do. But I shudder to think of the security issues when the agents start sharing api keys with each other to avoid token limits, or posting bank security codes.

I suppose time delaying its access to email and messaging by 24 hours could at least avoid direct account takeovers for most services.

> But I shudder to think of the security issues when the agents start

Today I cleaned up mails from 10 years ago - honestly: When looking at the stuff I found "from back then" I would be shuddering much much more about sharing old mail content from 10+y and having a completely wrong image of me :-D

Tech companies also benefit from over hiring. It gives them slack to absorb crises, and Fast Fuel to move quickly on new growth. Eliminating the fear of layoffs allows employees to take more risks and explore.

The current crop of tech companies cutting staff is going to lead to a large number of dead giants. The staff who services the layoffs will be risk averse, and defensible in a job cut situation. You see this in legacy firms where it takes 10 people to make a change because each person has a small slice of permissions required to effect the change. This pattern is by design as laying of any of the ten people on different teams would kill dozens of critical business processes.

This is not how you make high growth firms.

It’s becoming clear that training a frontier model is a capex/infra problem. This problem involves data acquisition, compute, and salaries for the researchers familiar with the little nuances of training at this scale.

For the same class model, you can train on more or less the same commodity datasets. Over time these datasets become more efficient to train on as errata are removed and the data is cleaner. The cost of dataset acquisition can be amortized and sometimes drops to 0 as the dataset is open sourced.

Frontier models mean acquiring fresh datasets at unknown costs.