GM! Welcome to Get Into AI!
Your favorite AI news guy back again here.
I'm that friend who hikes all the way up the hill to let you know if it’s worth your time!
Here is what I have for you:
AI Agent Autonomy is doubling every 7 months
Llama4 Rumors: Multimodal, 1M context window coming soon
"The model is the product": Big Labs changing the game
Alright, let’s dive in!
Three major headlines
Three main stories for the day.
1/ AI Agent Autonomy: Doubling Every 7 Months Since 2019
METR's new study answers the question we've all been asking: how fast are AI agents getting better at completing tasks without human help?
Turns out, pretty darn fast - doubling in capability every 7 months since GPT-2 in 2019.
They measured this with a clever metric: the "50%-task-completion time horizon". It’s basically how long would it take a human expert to complete tasks that current AI models can do with 50% success.
Right now, Claude 3.7 Sonnet can handle tasks that would take humans about 50 minutes.
At this rate, we'll have AIs with 1-day autonomy by 2028 and 1-month autonomy by late 2029.
Maybe I should start being nicer to my robot vacuum...
2/ Llama4 Rumors: Multimodal, 1M Context Window Coming Soon
The AI gossip mill is churning with rumors that Meta's Llama4 drops next month with multimodal capabilities and a massive 1 million token context window.
While the community is excited, there's healthy skepticism about whether such a long context is actually useful (or even works properly).
Models tend to get fuzzy memories long before hitting their theoretical limits.
There are also hopes Llama4 will be less censored than its predecessor, with developers eager for day-one support in projects like llama.cpp.
If the rumors are true, next month's going to be spicy for open-source AI.
3/ "The Model is the Product": Big Labs Changing the Game
A fascinating analysis argues we're entering a new era where "the model is the product" - not just the API or applications built on top of it.
The big labs (OpenAI, Anthropic, DeepSeek) are moving beyond just selling API access.
They're creating specialized models that handle entire workflows internally - like OpenAI's DeepResearch and Claude 3.7 - using specialized reinforcement learning to build in capabilities directly.
This shift could disrupt current API-dependent startups (awkward!) and change where value in the AI stack is captured.
As Naveen Rao of Databricks predicts, "all closed AI model providers will stop selling APIs in the next 2-3 years."
The winners might be companies that invest in training their own models rather than just wrapping existing ones.
It's like the difference between building your own house and decorating someone else's.
Catch you tomorrow! ✌️
That’s it for this week, folks! If you want more, be sure to follow our Twitter (@BarunBuilds)
🤝 Share Get Into AI with your friends!
Did you like today's issue? |
