New ask Hacker News story: Ask HN: Did LLMs Ruin APIs on the Web?
Ask HN: Did LLMs Ruin APIs on the Web?
3 by websap | 1 comments on Hacker News.
Saw this post on HN earlier today, https://ift.tt/Ge40BUA Separately the entire Reddit debacle of shutting down API access a few months ago, followed by the reports that Perplexity doesn't honor robots.txt, certain LLMs training on Youtube content, and Github Copilot being trained on opensource code without any opt-in. It just seems the tech is getting partitioned into a few different pieces in this space: 1. Data Custodians - Reddit, Yelp, etc that have a lot of organically generated data. 2. Model owners - OpenAI, Anthropic, Mistral, etc. that try to get licenses from data custodians to be able to make useful assistants. 3. Infra providers - AWS, Azure, Nvidia, AMD, etc. that plan to provide both of the cohorts above with compute. If I take this lens to the current crop of companies, it seems certain companies are really well positioned: 1. Apple - Apple has a near monopoly on the mobile device market. Their superior on-device processing stance, makes them a clear leader in delivering some of the best experiences with LLMs. I'd love to get a notification in my Messages if I try to make conflicting plans, or get summaries of group conversations, or summaries of emails. 2. Meta - They function as all 3 categories above. Own large amounts of data, train their own models, and do not rely on cloud providers, it would be interesting to see if they ever get into the chip manufacturing business. 3. Netflix - Similar to Meta, but more focussed in a narrow domain. 4. Google - It's astounding to me that Google hasn't already won this market. There must be some sort of crisis in Google leadership that they are unable to outpace and out innovate everyone. It almost seems like if you are not part of one of these larger companies, your best bet is to try and become a data custodian. Trying to build a user experience will become insanely hard given the exorbitant fees you'd be paying to infra providers and data custodians.
3 by websap | 1 comments on Hacker News.
Saw this post on HN earlier today, https://ift.tt/Ge40BUA Separately the entire Reddit debacle of shutting down API access a few months ago, followed by the reports that Perplexity doesn't honor robots.txt, certain LLMs training on Youtube content, and Github Copilot being trained on opensource code without any opt-in. It just seems the tech is getting partitioned into a few different pieces in this space: 1. Data Custodians - Reddit, Yelp, etc that have a lot of organically generated data. 2. Model owners - OpenAI, Anthropic, Mistral, etc. that try to get licenses from data custodians to be able to make useful assistants. 3. Infra providers - AWS, Azure, Nvidia, AMD, etc. that plan to provide both of the cohorts above with compute. If I take this lens to the current crop of companies, it seems certain companies are really well positioned: 1. Apple - Apple has a near monopoly on the mobile device market. Their superior on-device processing stance, makes them a clear leader in delivering some of the best experiences with LLMs. I'd love to get a notification in my Messages if I try to make conflicting plans, or get summaries of group conversations, or summaries of emails. 2. Meta - They function as all 3 categories above. Own large amounts of data, train their own models, and do not rely on cloud providers, it would be interesting to see if they ever get into the chip manufacturing business. 3. Netflix - Similar to Meta, but more focussed in a narrow domain. 4. Google - It's astounding to me that Google hasn't already won this market. There must be some sort of crisis in Google leadership that they are unable to outpace and out innovate everyone. It almost seems like if you are not part of one of these larger companies, your best bet is to try and become a data custodian. Trying to build a user experience will become insanely hard given the exorbitant fees you'd be paying to infra providers and data custodians.
Comments
Post a Comment