“________ AI News!” “Huge/Crazy/Major/Massive”. LLM Benchmark question, “Generate a YT thumbnail without using any of the words in this list.” (Love the channel content, great job!) 🙂
Already signs of greed are showing cause we dont yet have good open sourced models, they made the model that is suppose to be as cheap as possible 7x more expensive and now pushing regulations onto others to suppress creation of open source models just to make money instead of helping humanity advence faster
I think it's a great deal... Better performance than old 3.5 at a fraction of the price... For me personally the lower quality output of model is a no as I work on critical stuff programming and crunching numbers and complex function calling. Now I will be using haiku instead of relying 100% on models costing 15$ per million output token .. And use the higher models like 01 mini 01 preview and Claude 3.5 new for complex and more critical task.. This is great news Imo. Anthropic also still gives access to the older model if you want super cheap for non critical work.
7:28 This is a really good point. My concern is the down the road potentially we as customers won't have the ability to give our Gmail access to open AI. Maybe opening I needs to also create an email system
For me, searching through chat history is the biggest news! (10:18) I've wanted to do this for over a year, largely because I like giving an LLM a huge amount of context when I ask new questions.
"Maybe Microsoft is not giving [OpenAI] enough". From what I've heard, OpenAI keeps asking for more and Microsoft is making available all they can but it's not like they have that much more that would be sitting there idle. It's like complaining the grocery store is out of food after you've eaten all their stock *and* the reserves they had in the back. OpenAI have now started making deals with Oracle and even booked TSMC's *A16* process for manufacturing AI chips. A16 means 16 angstrom with 1 angstrom = 0.1 nanometer, so that's their new 1.6nm process… crazy times.
10:30 Hey Matt, You can absolutely chat with chat histories with memories just tell chat gpt to use memories only when replaying and it'll give you a chatable form where you can ask it personal things about yourself!!!
Great, interested in how bench are done! Today managed finally to extend openai 'swarm' to actually use assistants and multi-agent, in end thought better do self from scratch but got really good result at least on two documents from two different sources collated into one answer. THough think not big money there but interesting to see how works hands on. Long way to go seems but I need not do that thankfully.
Fully agree on Google and Apple still fully bing in the game on future dev. Apple have disappointed again but still have control of all the users' data and access to billions of apps that can be agentically connected in good time. Google are in a similar place, but Apple tend to deliver a more refined solution. The key thing will be once Apple decide to step into the commercial space with agentic apps and leverage user and company data for business actions. Same goes with Google in Workspace. Agentics is going to change so much.
Oh poor baby!😂 you should ask Matt for your money back oh wait it’s for free! so sick of leeches like you complaining when you get valuable information for free
Haiku seems like an odd choice. People don't choose small models for features, they choose them for price and accept the features they can get. So bumping the price up and justifying it with pricing seems odd.
Was waiting for your video, youtube feels empty empty
“________ AI News!” “Huge/Crazy/Major/Massive”. LLM Benchmark question, “Generate a YT thumbnail without using any of the words in this list.” (Love the channel content, great job!) 🙂
Do a deep dive on Regulation Matt, people deserve it!! 🔥🔥🔥🔥
Already signs of greed are showing cause we dont yet have good open sourced models, they made the model that is suppose to be as cheap as possible 7x more expensive and now pushing regulations onto others to suppress creation of open source models just to make money instead of helping humanity advence faster
I think it's a great deal... Better performance than old 3.5 at a fraction of the price... For me personally the lower quality output of model is a no as I work on critical stuff programming and crunching numbers and complex function calling. Now I will be using haiku instead of relying 100% on models costing 15$ per million output token
.. And use the higher models like 01 mini 01 preview and Claude 3.5 new for complex and more critical task.. This is great news Imo. Anthropic also still gives access to the older model if you want super cheap for non critical work.
7:28 This is a really good point. My concern is the down the road potentially we as customers won't have the ability to give our Gmail access to open AI. Maybe opening I needs to also create an email system
thx matthew b.
For me, searching through chat history is the biggest news! (10:18) I've wanted to do this for over a year, largely because I like giving an LLM a huge amount of context when I ask new questions.
"Maybe Microsoft is not giving [OpenAI] enough". From what I've heard, OpenAI keeps asking for more and Microsoft is making available all they can but it's not like they have that much more that would be sitting there idle. It's like complaining the grocery store is out of food after you've eaten all their stock *and* the reserves they had in the back. OpenAI have now started making deals with Oracle and even booked TSMC's *A16* process for manufacturing AI chips. A16 means 16 angstrom with 1 angstrom = 0.1 nanometer, so that's their new 1.6nm process… crazy times.
I don't think anyone realizes that the AI video game is bigger than we could ever imagine.
10:30 Hey Matt, You can absolutely chat with chat histories with memories just tell chat gpt to use memories only when replaying and it'll give you a chatable form where you can ask it personal things about yourself!!!
Great, interested in how bench are done! Today managed finally to extend openai 'swarm' to actually use assistants and multi-agent, in end thought better do self from scratch but got really good result at least on two documents from two different sources collated into one answer. THough think not big money there but interesting to see how works hands on. Long way to go seems but I need not do that thankfully.
Fully agree on Google and Apple still fully bing in the game on future dev. Apple have disappointed again but still have control of all the users' data and access to billions of apps that can be agentically connected in good time. Google are in a similar place, but Apple tend to deliver a more refined solution. The key thing will be once Apple decide to step into the commercial space with agentic apps and leverage user and company data for business actions. Same goes with Google in Workspace. Agentics is going to change so much.
I mean, why are people complaining? They get a better model than the previous 3.5 sonnet and its essentially cheaper. What are these guys smoking?
Great news
half of this video was an ad for vulture
Even though I know Matthew's videos are fine in spite of the clickbait, it still really turns me off and makes it much harder to click on his videos.
Thanks for clicking anyways!
Why do you say click bait ?
there's literally zero Clickbait in this title
Oh poor baby!😂 you should ask Matt for your money back oh wait it’s for free! so sick of leeches like you complaining when you get valuable information for free
Matt, you rock and you’re probably the best of any AI news on UA-cam. Keep up the good work and ignore the haters.
Why you havent made a video about Microsoft copilot studio latest talk?
AI news 😮
Sure Matt, I am sure you look forward to buy a robot with touch capabilities to "pet your dog" better LOL
Please make a video testing the new free Grok 2 API.
"Generated by AI", probably Claude 😂
Get a beard it suits you
Notwithstanding clickbait, or not, Looks like the Vultr link is broken... 😀
Haiku seems like an odd choice. People don't choose small models for features, they choose them for price and accept the features they can get. So bumping the price up and justifying it with pricing seems odd.
Anthropic is an incredibly dishonest company mark my words
HAVE A VIDEO THATS SPONSORED BY SOMETING OTHER THAN VULTR CHALLENGE (impossible)
They're just that good!
Just use multishot calls to llama and its still cheaper! Im using 8b!!!
👍
Claude 3.5 what a s,..t Ask a 500 lines of code Good luck 🎉😂