Really good visualizations overall.
100 karma · joined March 17, 2026
Really good visualizations overall.
These 'tests' are not labeled AGI by magic but because they are designed specificly for testing certain things a question answer test cant solve.
Gemini and OpenAI are at 80-90% at ARC-AGI-2 and its quite interesting to see the difference of challange between 2 and 3.
AGI progress means btw. general. So every additional dimension an agent can solve pushes that agent to be more general.
Nano Banana created a lot of noise.
But the reasoning of Gemini 3.1 Pro is really really good. Its hard to describe how good it became. I do not see the same quality from openai. Openai though is also super fast in response. A lot faster than just a few month ago.
For example: some german guy used the wrong word in describing an advantage of having a silencer and missuesd a word. Openai just said its nonsense, gemini suggested that its a typo and he wanted to write something else (gemini was correct).
It could also be that we are in a moat between "why is AGI not here yet" and "we need to build now the agentic platform stuff, that takes time".
Gemini pro is def slower than openai and I do not know if its because I use the pro version of gemini but not from openai. But it could also be that OpenAI has to work on subagents because Gemini def uses subagents and i was not able to find a source that OpenAI is doing this too.
Encoding more rules, more precise rules and alerting a human in case it thinks its off. Like salary increase by 20% gets flagged automatically. Revenue drop bey x % too.
It could even go so far that the maker of these systems will insure you for their use.
It just needs to be cheaper than all the humans in the loop and if you train it once, you can copy it unlimited time. Scaling effect of software for tasks we need to train a human again and again.
It could also be agent systems which do this. Like a company building and designing the HR USA Healthcare agent specialized in SAP HR. Another one for HR Brazil Healthcare agent specialized in another HR software.
Humans are really expensive and you have to train them regularly and every single on of them.
So a good PM running 1-3 teams, will only need 1-3 agentic ai teams instead.
You will have an agent like your seo expert, this agent will be able to use common tools like google seo, facebook seo etc. and you will teach how you want it to do its 'job'.
You will have a way of delivering your requirements to it, it will run in the background, might ask for feedback but will otherwise do stuff similiar to whatever person was doing it before.
There might be some transition phase like verifing the data of the real person vs. the agentic ai then moving over to only validation until the agentic agent is in avg as good as a human. Then the human will be gone.
Agentic will take basic support tasks (its actually already doing this) first, then more complicated things etc.
For this we need an ecosystem aka the agentic ai platform, interconnect between agent and tools and this stuff is currently getting build by someone one way or the other.
On scale we need more capacity and these agents will also cost more money than a 20$ subscription.
But if you have a, lets say SAP agent, it will be build once, trained once and than used by everyone. Instead of a person using a HR system or billing system, the agent will bridge the gap between data and system.
So they need to be able to do image generation, for which they need image data. They also need to be able to analyze videos for more and better training data like learning or teaching there models from yt and other sources.
So they have image generation, image dataset and video dataset. Its not far fetched ata ll or desperate to leverage this base for playing around with video generation.
And despite how much money they burn, for a company that size, trying out video generation wasn't that high of a goal post.
I'm really surprised by there move and can only imagine that the progress of other models from google and antrophic pulls their teeth and no longer want to invest the compute (not money) to leverage their compute for their main models.
Having Disney on their side was def quite a smart/interesting move.
At least from one interview, they def had resource issues last year and teams had to fight for it. Can easily be that sora was always priortized down and they realized it doesn't make sense to spend that much capacity while then not being able to push their main model.
AI/LLMs are not getting stupid in something they are already good in (or very seldom if).
So i see it more as an evolving thing not a wicked thing.
Newest claude release made context better, for this information I already need to know claude and know what a context is and what the problem with context length is.
I'm following AI for a few years now and def not have the feeling at all its wicked.
For example: gemini became a lot better in a lot more tasks. How do I know? because i also have very basic benchmarks or lets say "things which haven't worked" are my benchmark.
Without playing around with it, you wouldn't know when to use an LLM and when not.
But AI is a beast.
Its A LOT to learn. RAG, LLMs, Architecture, tooling, ecosystem, frameworks, approaches, terms etc. and this will not go away.
Its clear today already and it was clear with GPT-3 that this is the next thing and in comparison to other 'next things' its the next thing in the perfect environment: The internet allows for fast communication and we never have been as fast and flexible and global scaled manufactoring than today.
Which means whatever the internet killed and changed, will happen / is happening a lot faster with ai.
And tbh. if someone gets fired in the AI future, it will always be the person who knows less about AI and knows less about how to leverage than the other person.
For me personally, i just enjoy the whole new frontier of approaches, technologies and progress.
But i would recommend EVERYONE to regularly spend time with this technology. Play around regularly. You don't need to use it but you will not gain any gut knowledge of models vs. models and it will be A LOT to learn when it crosses the the line for whatever you do.