Good question. Thanks for asking. I’m not sure what I’d like to see is easy with today’s models. Voice summary & associated conversation generation tends to be fairly bland. Many guard rails. Grok shows a terrible fix. Still, I’m interested in sharpening as a path to summarizing. I think that will require a different model spec in post training. 18 months ago I explored sharpening with prompts, but was dissatisfied. I think I might do much better with today’s models & agents. You may be asking about what new types of apps I’d like to see from EV. I’d like a EV LCP server in paid and unpaid versions. Unpaid would have results limited to unpaid EV. Paid would include cleaned up transcriptions of audio and video content. (Ideally it would be based on voice to text fine tuned to handle technical terms well.) there could be a third tier: that would include references in the perplexity style. That would probably require a partnership with Perplexity or similar.
Thanks this was a reminder that energy capacity is a limiting factor for inference compute; though, energy efficiency innovations could decouple these two…
Getting this Ev update also as quick morning podcast briefing would also be great! Nothing fancy, just let me listen while driving to work, walking with the dog or cooking lunch.
You can listen to it in the Substack app (see the play button at the top of the post). We’ll consider making it available on the podcast feed as well, nice idea!
I currently spend about 150 minutes a day shopping and cooking. Can you seed me a cook/shopper to free up time for your app of the week?
I had a vague idea that something like this wa possible, but the concrete example is thrilling.
What would you like the app to help you do?
Good question. Thanks for asking. I’m not sure what I’d like to see is easy with today’s models. Voice summary & associated conversation generation tends to be fairly bland. Many guard rails. Grok shows a terrible fix. Still, I’m interested in sharpening as a path to summarizing. I think that will require a different model spec in post training. 18 months ago I explored sharpening with prompts, but was dissatisfied. I think I might do much better with today’s models & agents. You may be asking about what new types of apps I’d like to see from EV. I’d like a EV LCP server in paid and unpaid versions. Unpaid would have results limited to unpaid EV. Paid would include cleaned up transcriptions of audio and video content. (Ideally it would be based on voice to text fine tuned to handle technical terms well.) there could be a third tier: that would include references in the perplexity style. That would probably require a partnership with Perplexity or similar.
Thanks this was a reminder that energy capacity is a limiting factor for inference compute; though, energy efficiency innovations could decouple these two…
Getting this Ev update also as quick morning podcast briefing would also be great! Nothing fancy, just let me listen while driving to work, walking with the dog or cooking lunch.
You can listen to it in the Substack app (see the play button at the top of the post). We’ll consider making it available on the podcast feed as well, nice idea!