will to code
305 subscribers
355 photos
63 videos
21 files
152 links
Download Telegram
Neurons that fire together wire together
If I was an investor and had the opportunity I would happily invest 100% of my portfolio into major ai companies (open ai, anthropic, groq, moonshot...) even at this stage

The sentiment right now is that they are bubbles about to implode

My gut tells me otherwise
Karpathy is soooo right

> How good is your planning?
> How good is your short/long term memory?
> How much do you care?

Imagine creating a detailed memory palace on what you want to get done and constantly updating it... Literally GOD MODE
(...also being able to recall past sessions 🫠)

The landscape has completely changed

Those niche skills matter more than anything now
We need to bring back blimps

Its one of the most exciting modes of transport to come out of a dynamic era

Think about all the advancements in AI and nano tech...

Autonomy, Efficiency and very hard to crash... Its better than telling people to go on autonomous planes which are practically cruise missiles.
The hardest thing in the world is simple
💯3
I strongly believe smart context management would positively impact our AI workflows

> You tell AI write some code
> It writes code
> Code has error
> You tell it to fix the error
> It fixes it

But now that "fixing" is in the context and sitting there not contributing to anything. What we need is this:

> It goes back to the earlier code it sent
> Adds the fix there
> Deletes the messages of you telling it to fix the code

We don't know anything at this point. What if it just follows the pattern of making a mistake and you telling it to fix it.

Or like Antrhopics recent research what if it just takes on the personality of a sloppy dev... What if there was an agent which goes through the context and makes sure that the LLM's 'personality' is competent.

The context manager decides what the long and short term memory needs to be.

Dynamic context. Kinda like us.
Evolution is nature's way of finding product market fit

- Charles Darwin
1
AI is wayyyyyy under hyped

I was one of the people on twitter saying gpt-5 will be AGI in the early days of gpt-4: which is now considered an unusable model by any standards

At this moment, we are the bottle neck: Skill Issue

If you look at nations, or groups of people in general, the consensus always happens—or rather, policies always get enacted—after the fact. No action is taken during a decline in birthrate. For a better example: when the birth rate goes from 4 to 3, nobody sounds the alarm. It’s only when it reaches 1 or a negative number that policies start to get passed and the general public agrees that their birthrate is declining.

In the same way nobody is aware of the insane performance leap from gpt-4 to gpt-5.4
> You just need to extrapolate a bit to understand the situation we are in

You have to ignore the consensus of "ai slop" and try to push these models to any limit you can imagine them having
👍5👎4
German word for library:

Bibliotek 🗿
New open source drop!!!!
> 6x cheaper than opus
🔥3
It costs $25 input $125 output 💀
The pentagon really fumbled this one

Not good when AGI is your supply chain risk lol
*meta has joined the chat