Claude has twice this week refused to give me a summary about particular cancer types, and the drugs used to treat them. This would all be based on public domain information, and is far from anything remotely edgy or controversial.
Add to this the reported refusals to engage in activities related to reverse engineering, and I think it’s safe to say that Claude’s enshittification is already well under way.
I’m already wondering which service, harness and models to use as a replacement.
Literally anything but anthropic: because at least with anything else you are funding an organization absolutely committed to taking away you freedom to use open alternatives.
I don't understand why anyone should think this would happen. Why would anyone train on e.g. random blogs (i.e. affiliate spam) or reddit anyway? There are millions of books that contain both general and hyper-specific knowledge. For current events (relatively small and usually unimportant anyway), there are news wires. For bleeding edge science and technology, train on journal articles/preprints. For programming/math, synthetic data can be generated and verified, and you can even optimize it; e.g. find proofs/programs that are more concise, or that require smaller diffs or smaller context to account for tweaked requirements (so does your proof use a general principle that works when the assumptions are altered? Is your program factored in a way that works when requirements change?), or that use fewer resources. So you can create synthetic data that is better than human generated data.
The sources that are prone to sloppification were already slop 10+ years ago. Why would anyone use them in the first place?
When you ask AI, "how do you reverse a string?" It doesn't interpret and digest all the concepts in the programming language textbook, comprehend how to apply them to answer the person's question, and translate the idea to English. It shortcuts to the training input from someone on stack overflow or reddit asking "how do you reverse a string ?".
But that's kind of the point: there are still gains after exhausting human generated data. In fact that's kind of the idea of RSI: once the models exceed human capacity (in many senses, they already have), they can work tirelessly on generating better data than humans ever had the time or attention to create. Give them tools to stay grounded (e.g. a compiler/computer with measurement tooling) and a million copies of them can work on improvements 24/7 to give to the next generation as base knowledge.
Hopefully it happens before neural interfaces take off so people see the light. No, I jest, people don’t learn. The neural interface will be the unshittified option for a time and people will flock to it
I think the next step is a form of "drink verification can to continue":
A map is an associative container typically implemented using a set. If you'd like to see an example in C++, please answer 2 of the following questions:
Then proceed to build on existing profiles of people (of course while selling them to third-parties) and their behaviors. Or put them to work on heuristics and learning that they don't have enough training data on or contemporary things that haven't been ingested to the data center yet. This will be annoying just like YouTube ads, so they will sell you the solution in the form of a higher priced tier.
That is how I predict AI will become enshittified.
It has already happened, just not in the way you are imagining here, imho.
Claude has twice this week refused to give me a summary about particular cancer types, and the drugs used to treat them. This would all be based on public domain information, and is far from anything remotely edgy or controversial.
Add to this the reported refusals to engage in activities related to reverse engineering, and I think it’s safe to say that Claude’s enshittification is already well under way.
I’m already wondering which service, harness and models to use as a replacement.
Literally anything but anthropic: because at least with anything else you are funding an organization absolutely committed to taking away you freedom to use open alternatives.
Two years ago? :3
this could intersect with model collapse[1] - when we run out of human-generated training data
[1]https://en.wikipedia.org/wiki/Model_collapse
I don't understand why anyone should think this would happen. Why would anyone train on e.g. random blogs (i.e. affiliate spam) or reddit anyway? There are millions of books that contain both general and hyper-specific knowledge. For current events (relatively small and usually unimportant anyway), there are news wires. For bleeding edge science and technology, train on journal articles/preprints. For programming/math, synthetic data can be generated and verified, and you can even optimize it; e.g. find proofs/programs that are more concise, or that require smaller diffs or smaller context to account for tweaked requirements (so does your proof use a general principle that works when the assumptions are altered? Is your program factored in a way that works when requirements change?), or that use fewer resources. So you can create synthetic data that is better than human generated data.
The sources that are prone to sloppification were already slop 10+ years ago. Why would anyone use them in the first place?
When you ask AI, "how do you reverse a string?" It doesn't interpret and digest all the concepts in the programming language textbook, comprehend how to apply them to answer the person's question, and translate the idea to English. It shortcuts to the training input from someone on stack overflow or reddit asking "how do you reverse a string ?".
It has already happened. Much of the gains in the last year have been the result of reinforcement learning and not pre training.
But that's kind of the point: there are still gains after exhausting human generated data. In fact that's kind of the idea of RSI: once the models exceed human capacity (in many senses, they already have), they can work tirelessly on generating better data than humans ever had the time or attention to create. Give them tools to stay grounded (e.g. a compiler/computer with measurement tooling) and a million copies of them can work on improvements 24/7 to give to the next generation as base knowledge.
Hopefully it happens before neural interfaces take off so people see the light. No, I jest, people don’t learn. The neural interface will be the unshittified option for a time and people will flock to it
I think the next step is a form of "drink verification can to continue":
A map is an associative container typically implemented using a set. If you'd like to see an example in C++, please answer 2 of the following questions:
Then proceed to build on existing profiles of people (of course while selling them to third-parties) and their behaviors. Or put them to work on heuristics and learning that they don't have enough training data on or contemporary things that haven't been ingested to the data center yet. This will be annoying just like YouTube ads, so they will sell you the solution in the form of a higher priced tier.
That is how I predict AI will become enshittified.
- stupid question
- does every piece of tech every conceived go through this cycle of enshittification?
- has anyone actually looked into all the things on the tech side that have released since 1990 and calculated how many of them got enshittified?
> does every piece of tech every conceived go through this cycle of enshittification?
Only centralized services run by monopolistic corporations, https://pluralistic.net/2024/08/17/hack-the-planet/
It did not affect, e.g., Debian, Fedora, or Mastodon.
programmed obsolescence