@ai4you.bsky.social
I love complex artificial intelligence systems. Every day a new surprise. Every day up to 800 new research pre-prints to design new systems, from medical to financial to social AI. I do believe, that science should be open to everybody.
DeepSeek has released a new version of R1, called R1 0528.
No R2, but an improved R1.
I already tested it on complex reasoning, and this open source model is impressive.
youtu.be/toailYMTAKo?...
New video on creating a new app - for my AI research papers.
No cursor, no windsurf, no bolt, no lovable, no replit.
Google has a new service: try it for free in AI Studio.
Code your app in 15 minutes, from scratch:
youtu.be/x7zrS6xXmgM?...
New Qwen 3 30B MoE A3B model is really interesting.
I performed a cascading logic test on it, in thinking and non- thinking mode.
Live recording
youtu.be/u-WXyeV1tsw?...
Current AI methods like SFT or RL are not optimized for quantum computing.
New Quantum AI Framework has been published. AI comes closer to theoretical physics.
Further insights here
youtu.be/vIb3B0PIklE?...
If you want to learn HOW to CODE agent and multi-agent systems, today is perfect time to start.
Google released a new Agent framework: ADK. Maybe the most efficient way to code agents with tools, like internet search, MCP and so much more.
youtu.be/Geo8LzCHoMQ?...
META AI is in trouble.
The REAL Llama 4 models fail to perform as marketed.
"Optimized" Llama 4 for Benchmark Tests only.
#llama4
youtube.com/post/UgkxyO3...
New Llama 4 Maverick 400B model, official benchmark:
youtube.com/post/UgkxWRX...
I am really interested in the new AI research by DeepSeek.
They published a new Reward model for their next Reasoning model:
Explained in detail here
youtu.be/9KMxNZ2CvUg?...
Llama 4 Maverick 400B MoE
Tested on logic and causal reasoning: press a sequence of 5 elevator buttons to go up.
Llama 4 400B failed ... 6 times .... and got frustrated with me!
Live recording of Llama 4 performance
youtu.be/8G-GI4bvWZU?...
First real world TEST
Llama 4 Maverick 400B
On causal reasoning
#llama4
New video
youtu.be/12lAM-xPvu8?...
New DeepSeek V3 0324 has an impressive performance, but is not a reasoning LLM.
So I manually build a Single Task R2 system - out of DeepSeek V3 0324.
I show you all the steps for a single Task R2 - if you want to upgrade already.
youtu.be/TxtSD8DDqKk?...
New small DeepSeek R1 Models.
New 14B and 32B Light-R1 models for local use. Open- source.
youtu.be/FAg4v2xaLYc?...
Combine your local LLM with a cloud based reasoning #LLM.
New protocol to run your Ollama DeepSeek #R1 locally and only use #Sonnet 3.7 for the heavy thinking.
#Stanford Univ open sourced a new code
youtu.be/L-WfRaSPE2A?...
The old dream: Ai will do research for us.
Google's new Co-scientist with 7 interacting special AI agents is the latest try.
But be careful: to automate scientific research with AI can have massive side effects.
New video
youtu.be/TUo1VeeBgOU?...
Maybe Grok 3 or Sonnet 3.7 is not a good choice for your AI agents.
I will show you a new optimization where given a task, multi LLMs will be selected in a multi agent config - according to their performance. Not their Hype.
New video:
youtu.be/7HxDU8K59k8?...
Stanford Univ innovated tool use for AI agents.
Since agents can integrate other specialized AI agents as tools, a new method for multi tool use was invented - without the need to train any Model or supervisor agent.
New video
youtu.be/4828sGfx7dk?...
Grok 3 is now free for 10 queries per day
First tests w/ Grok 3 THINK (the deep reasoning mode) - similar to DeepSeek R1 reasoning
youtu.be/1trUPXnREmA?...
Hi community,
Today Perplexity.AI gave us their Deep Research Engine. Free Deep Research!!!
Instead of OpenAI's $200 option or Google's advanced option, Perplexity offers 5 free runs of Deep Research per day.
I did live testing - regarding the latest AI research topics:
youtu.be/Z9IpO3TTskU?...
AI Clones of human individuals are rather easy to code.
Next level is to code an AI Agent with the personal values, individual characteristics and private thoughts of an individual.
Deep reasoning is the way to explore those - for an AI Agent representing YOU
New video
youtu.be/gnJqsO8Mm1w?...
If you want to see a product presentation by OpenAI, that is really special, why not have a look at OpenAI's homepage for the new product: Deep Research.
I compare the performance of the new Deep Search to a human and to a vanilla ChatGPT (free version).
youtu.be/tLnZBUuxNAI?...
There is a new OPEN R1 initiative.
Yes, DeepSeek R1 is open- source, but some secrets still remain.
Open R1 is a new effort by the open-source community, to uncover the complete complexity of the latest AI.
More details and how you can interact:
youtu.be/2ENvGkkK36E?...
Improved AI reasoning with knowledge graphs and multi agent systems.
Improve on your Knowledge Graphs in GraphRAG.
The idea is simple: instead of planning your path node by node, calculate community to community.
Faster, cheaper and more efficient:
youtu.be/DoI4nWQuywI?...
Fact checking is essential.
Especially for AI systems.
To reduce AI hallucinations, new research on AI internal fact checking has been published.
The ultimate AI Fact checking method? It comes from medical record fact checking!
This video explains it:
youtu.be/ry3R7k6x1Pg?...
Can we really learn from LLMs? Can they become our learning engines?
I am perform a live test: OpenAI on vs DeepSeek R1 on learning and explaining new AI research topics.
What do you think? Is it worth paying triple prices for o1?
Here a direct comparison of o1 and R1
youtu.be/HM92mmG6YTs?...
A performance comparison of new #Gemini Thinking 01-21 LLM (published today) and DeepSeek R1 (published yesterday). #R1 #Reasoning
If you want to see a frustrated LLM that went into deep #CoT on my reasoning task and declares: I GIVE UP! ...have a look at my new video:
youtu.be/jb6egub3JDk?...
DeepSeek published new open-sourced Reasoning models (R1).
Including small Language Models distilled from DeepSeek R1 to Qwen 32B down to Qwen 1.5B SLM.
All new models explained:
youtu.be/KhY9XK1jGCQ?...
Google released prototypes of a new Transformer model with RNN memories.
For a 4 million token context length. Self-attention with RNN linear compute complexity memory.
Called TITANS, their LLM architecture is quite challenging:
youtu.be/X2GpzYfy_sE?...
Do you remember Grokking?
The late onset of the performance phase of LLMs?
(Note: not related to the Grok named models).
Now finally we have an explanation why Grokking happens in LLMs /Transformers:
youtu.be/SRfJQews1AU?...
AI can have real positive effects in research: like in protein engineering.
From proteins to enzymes, the development of new biochemical compounds is perfect for AI support in biotech applications.
The latest AI models for protein design? You find it here:
youtu.be/9cNxgYhmAAg?...
System 2 Reasoning - like o1 - on your LLM?
Why not! Inference compute ot Test-Time Compute are new performance vectors in AI.
An open-source training script w/ training data to update your LLM to o1 reasoning. 🧪 #AI
Sky-T1 by UC Berkeley open-sourced everything.
youtu.be/ZmliPzGENMM?...