EP76: Can AI Fix Its Own Mistakes? (Reflection 70B) & How Much Will You Pay for AI Productivity?

----
CHAPTERS:
00:00 - Days of AI Models Lives
04:02 - Reflection 70B Open Source Model: Is It The Best Open Source AI Model or Just Great Prompt Engineering?
24:48 - Is Microsoft CoPilot a Dud? What Actually Makes you More Productive in Enterprise AI.
36:15 - OpenAI Floats $2,000/month for New Models Strawberry (Q*) and Orion. Is it Expensive or Cheap for Potential Gains?
55:51 - Boom Factor for Reflection 70B & Final Thoughts
-----
Thanks for listening and all of your support of the show!

[tunesafari]

It’s officially Friday now, thanks gents, great episode.

[ozzukee]

Great, yet another perfectly average episode

[FredPauling]

Listening to the podcast and then jumping into Sim theory to try out new models and functionality is so good.

[micbab-vgmu]

Copilot 365: Slower than a sloth on dial-up, with more bugs than a picnic in July.

[DodefiEnglish]

Thanks. Great show again. Love Sim Theory 🙂

[Limitless]

Reflection 70B is (in effect) the first step in recursive learning for Ai. It is level 1 of synthetic data recursion.

[WarClonk]

Its really unfortunate that you guys didnt get Reflection to run. I know it is probably more hype than red meat but it can also only be really dismissed when someone tested it thorougly. It is totally possible that there are some "hacks" that few have found yet. And lets not forget we are talking about a 70b model here. So while I am skeptical, I am still intrigued to see the 405b model

[joser]

there are now some Q4 versions for Reflection that work well on a laptop... (I tried with both Ollama and LMStudio) both quite ok, a bit slow but as usual with these larger models... Looks OK, not like Claude 3.5 sonnet or GPT-4o, sure not, but for this size looks better than others of that size

[DefaultModeNetworkAI]

Were using gpt4o with over 40 different API tools it can call to retrieve data from various locations throughout our company with the intent of any type of question one of our thousands of employees ask they get the data retrieved from our company data. Could be project info, HR employee data, marketing data, etc. Both unstructured content and structured DB content. Works amazingly well and is providing very high ROI. I believe each company needs to really figure out what their business needs are and what use cases they want to solve. Too many just try to force AI into their workflows without having a real need. Its just another tool to use, solve problems dont just use a "tech" just to say your company "has" or "uses" AI. My 2 pennies :)

[isaklytting]

48:10 Wait, what software is it that can do this?? Is this something I can add to Visual Studio?

[adotinthecosmos]

Great discussion. You guys are absolutely right, we don't want these segmented AI's embedded within applications.. we just want to have a well rounded AI that we can ping with a quick question. The one place that I see specialized AI taking over is inside of applications with clunky or complicated UI's. These applications will have a bunch of user's onboard if the functionality is now just a small conversation away.

[augmentos]

Can you guy do an update on where Mamba and Bitnet 1.58 architectures are at? They showed early promise, what’s the latest on them??

[glennmontague]

Isn't the scratchpad idea something that has been in Anthropic's guide to prompt engineering for a couple of months?

[breaktherules]

is there a website I could chat with reflection ai for free without downloading?

[mangoauto]

They were like teenagers with a unfounded fear of ‘getting caught’ by their parents (higher ups) for really nothing of significance

[martinseal]

Yh it still has no understanding of what you asked it and having to check the prompt will just cost more, nothing revolutionary same technology

[FutureFictionAcademy]

Really enjoyed the podcast today guys. Great episode, caught it while eating breakfast. Gave you a shout out on my live writing today (we're all authors working with AI to write books). Excited to check out your other episodes. -EAW

[KristopherRichards]

My takeaway is that synthetic data will be more valuable than real-world data. Also, this is why the true system prompt is so secretive.

[thetrolle]

A mistake in the title (it's/its), when the title is itself about mistakes. Other content good, though - so, on balance.... average.

[christianrusso]

People who use Ai are falling behind to people who avoid Ai. AI is nothing but rubbish and that’s coming from someone who’s very tech savvy and skilled with computers.