gcollard-

I’d also be interested in seeing the rate of hallucinated citations from the pre-ChatGPT era.

I’m not entirely convinced this is purely an LLM-related issue. I’ve definitely come across countless misattributed citations in big four reports long before generative AI became widespread.

Testing this hardware LLM (LLAMA 3.1 8B on a chip) I get ~16k tokens per second.

With frontier models plateauing, I’ve been convinced AI will end up like bitcoin mining, and that NVIDIA’s general-purpose GPUs will be replaced by model-specific chips.

Glad to see someone innovating in this space.

Forget everything you know and consider that it might be a misguided and risky negotiation tactic.

Disclaimer: This is not business advice and should be read using Cartman’s voice.

Step 1: Announce publicly that you are not renewing your contract.

Step 2: If the market has viable alternatives or the service you are negotiating isn’t that hard to replicate, other actors will manifest to fill in the gaps, especially if your business is attractive. (E.g., The top comment is building an alternative; other comments point to alternative services.)

Step 3: Congratulations, you now have leverage for a significant discount with your previous provider because they face the real prospect of losing your business entirely to a competitor. If the competitor is private, you can even double dip by investing in their company before attributing them the contract.

All the publicly accessible sources you mentioned have already been scraped or licensed to avoid legal issues. This is why it’s often said, “there’s no public data left to train on.”

For evidence of this, consider observing non-English-speaking young children (ages 2–6) using ChatGPT’s voice mode. The multimodal model frequently interprets a significant portion of their speech as “thank you for watching my video,” reflecting child-like patterns learned from YouTube videos.

Subdomains.

Parent is working on scheduling tools for NA meetings. Yes some things are important and can’t be broken without some serious real world consequences.

MVP is “the first thing that stick”. The method to get there should be meticulously engineered to cut as much iteration time between prototypes as possible. When your ugly prototype shows strong traction by its own and test users are asking you to keep the prototype then you know you have a rough diamond to cut.

Obviously when you have a sizeable user base, know the market very well and your market isn’t that much competitive you are at an advantage.

Beware of market expertise ego and dogfooding as it only reinforce your expertise and biases. Bring industry newcomers to your project for feedback, it’s the only way to make sure your expertise doesn’t make your product only valuable for yourself.

There’s rights and licensing fees associated to it.

In a couple years this might be displayed in a museum somewhere.

Modern art like monochromes are still displayed to this day.

Markdown readmes on github.

« Mui tells TechCrunch that startups must fulfill two criteria in order to automatically invest: first, the co-founder must be a member of the class, and second, they must raise a round of $750,000 or more from a reputable institutional investor »

I think they just don’t pick winners themselves