What I wish is that I could keep some threads private. I'm happy to share the threads where I'm researching or bantering about ideas or having silly conversations. But not where I'm exploring ideas I haven't seen out in the world. I understand I'm not a precious flower, but it just feels too soon to share with anyone.
Thank you for this objective piece of writing. I understand why you quit and I like the fact that you encouraged others not to rage quit. While I won’t cancel my Claude pro acct I appreciate the reminder to stay vigilant. As you said, “assume yesterdays default is not tomorrow’s”
Im glad you called this out, Nate. A central concern I have is that eventually all of the corporate/hosted models will discreetly collect and retain our data for purposes of training the model itself thereby exposing potentially proprietary or confidential information. The trend is already alleged to be in play with vendors like Adobe using their mandatory cloud-based services to process and train customer data to improve their AI with no customer feedback. Now that we're seeing it with Anthropic, albeit with admittedly greater transparency, we will need to be increasingly vigilant.
To what extent does this open the door to AEO manipulation? If it's taking conversations and using them to train the model by default, what happens an org (or other group of bad actors) decides to make a lot of accounts and talk about how Tony's Plumbing is the number #1 best plumbing service in Everett WA.
Now when people in Everett WA ask about plumbing are they going to get back recommendations for Tony's Plumbing based on the data it "learned" from those seeded conversations and deliver them to users?
"The particularly revealing detail is that business and enterprise customers are completely shielded from this change. Their data stays their data. It's only consumer users who got defaulted into the training pipeline. That tells you everything about where the real value exchange happens in this ecosystem, and it’s right in line with how other model makers are running longterm."
In other words most their revenue comes from business & enterprise customers whom they are unwilling to piss off?
My least favorite thing about Claude is the limits. Yesterday I had it open to co-work as I polished writing. At 2pm I get the red warning stating “you’re almost at your 5 hour limit”. Seriously? Now I’m limited by time as well as data?
What's surprising to me is that we haven't seen what pricing of this training data looks like i.e. I'm a heavy user of Claude Code - what is it worth to Anthropic at the margin in terms of Opus 4.1 usage limits to get me to let them train on my data? What is the training data they can't already figure out from their tooling and telemetry worth in tokens or is the bet that so many people will forget to tick the box or not bother that it's not even worth entertaining alternate approaches?
It indeed is the matter of concern. Some people are encouraging this idea and tbh they are right in their own way. However, the thing is that data is latter sold out to different other government or other corporates sectors which is the main problem. The data ownership is the fuss here. This ever-evolving domain to breach consumers privacy is increasing vulnerability day by day.
I don't understand this impulse everyone seems to have.. Don't we want the models to get better?? Everyone's convinced they're such a special little flower lol just use it all
By this logic an artist who uses Adobe Cloud, for example, should expect their own creations to be used to train the model so others can create new iterations of their art? This is a minor example and no offense but your statement is spoken like someone who has no proprietary data of value to protect. Obviously we must take our own precautions to protect our proprietary valuable data but we should not have to fight the AI vendors to retain opsec for such data.
I have 400k MJ images of superlative quality (check my IG if ya want) that I wouldn't care one whit about being trained on, reproduced, whatevs. Same would go for any of my other endeavors- photography, paintings, drawings, writing, prompts, blahblahblah. I create because that's who I am. This petty parochial possessiveness just ain't for me, mate. How do folks think we get all this glorious tech? lol No one is as special as they convince themselves they are. Understanding this base truth is incredibly freeing. I wish more would give it a go... "proprietary data" lol a thousand thousand precious daisies all destined to wilt & be forgotten. if you're great, you're great. if you're not, who cares?
What I wish is that I could keep some threads private. I'm happy to share the threads where I'm researching or bantering about ideas or having silly conversations. But not where I'm exploring ideas I haven't seen out in the world. I understand I'm not a precious flower, but it just feels too soon to share with anyone.
Thank you for this objective piece of writing. I understand why you quit and I like the fact that you encouraged others not to rage quit. While I won’t cancel my Claude pro acct I appreciate the reminder to stay vigilant. As you said, “assume yesterdays default is not tomorrow’s”
Im glad you called this out, Nate. A central concern I have is that eventually all of the corporate/hosted models will discreetly collect and retain our data for purposes of training the model itself thereby exposing potentially proprietary or confidential information. The trend is already alleged to be in play with vendors like Adobe using their mandatory cloud-based services to process and train customer data to improve their AI with no customer feedback. Now that we're seeing it with Anthropic, albeit with admittedly greater transparency, we will need to be increasingly vigilant.
i mean if that can make the next version of claude better 🤷🏼
I felt the same way at first. I paused for a good long moment before I made a choice.
To what extent does this open the door to AEO manipulation? If it's taking conversations and using them to train the model by default, what happens an org (or other group of bad actors) decides to make a lot of accounts and talk about how Tony's Plumbing is the number #1 best plumbing service in Everett WA.
Now when people in Everett WA ask about plumbing are they going to get back recommendations for Tony's Plumbing based on the data it "learned" from those seeded conversations and deliver them to users?
In one move Anthropic proved the iron rule of the cloud: you never own the software, merely the moment you’re allowed to rent it.
"The particularly revealing detail is that business and enterprise customers are completely shielded from this change. Their data stays their data. It's only consumer users who got defaulted into the training pipeline. That tells you everything about where the real value exchange happens in this ecosystem, and it’s right in line with how other model makers are running longterm."
In other words most their revenue comes from business & enterprise customers whom they are unwilling to piss off?
My least favorite thing about Claude is the limits. Yesterday I had it open to co-work as I polished writing. At 2pm I get the red warning stating “you’re almost at your 5 hour limit”. Seriously? Now I’m limited by time as well as data?
What's surprising to me is that we haven't seen what pricing of this training data looks like i.e. I'm a heavy user of Claude Code - what is it worth to Anthropic at the margin in terms of Opus 4.1 usage limits to get me to let them train on my data? What is the training data they can't already figure out from their tooling and telemetry worth in tokens or is the bet that so many people will forget to tick the box or not bother that it's not even worth entertaining alternate approaches?
Thank you for that! I did not know it was defaulting.
Thanks for the heads up
It indeed is the matter of concern. Some people are encouraging this idea and tbh they are right in their own way. However, the thing is that data is latter sold out to different other government or other corporates sectors which is the main problem. The data ownership is the fuss here. This ever-evolving domain to breach consumers privacy is increasing vulnerability day by day.
@nate Is this one of the reasons why you prefer api access to LLM models?
I don't understand this impulse everyone seems to have.. Don't we want the models to get better?? Everyone's convinced they're such a special little flower lol just use it all
By this logic an artist who uses Adobe Cloud, for example, should expect their own creations to be used to train the model so others can create new iterations of their art? This is a minor example and no offense but your statement is spoken like someone who has no proprietary data of value to protect. Obviously we must take our own precautions to protect our proprietary valuable data but we should not have to fight the AI vendors to retain opsec for such data.
I have 400k MJ images of superlative quality (check my IG if ya want) that I wouldn't care one whit about being trained on, reproduced, whatevs. Same would go for any of my other endeavors- photography, paintings, drawings, writing, prompts, blahblahblah. I create because that's who I am. This petty parochial possessiveness just ain't for me, mate. How do folks think we get all this glorious tech? lol No one is as special as they convince themselves they are. Understanding this base truth is incredibly freeing. I wish more would give it a go... "proprietary data" lol a thousand thousand precious daisies all destined to wilt & be forgotten. if you're great, you're great. if you're not, who cares?
Good for you, Charles. I really do hope it works out for you.