ai4you.bsky.social - Profile | ThreadSky | a Reddit-style client for Bluesky

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?...

submitted 2 days ago • 0 comments

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?...

submitted 7 days ago • 0 comments

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?...

submitted 7 days ago • 0 comments

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?...

submitted 15 days ago • 0 comments

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?...

submitted 24 days ago • 0 comments

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?...

submitted 24 days ago • 0 comments

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?...

submitted 24 days ago • 0 comments

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?...

submitted 30 days ago • 0 comments

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?...

submitted 31 days ago • 1 comment

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?...

submitted 32 days ago • 0 comments

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?...

submitted 34 days ago • 0 comments

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?...

submitted 38 days ago • 0 comments

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?...

submitted 40 days ago • 0 comments

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?...

submitted 41 days ago • 0 comments

Can small Language models with iterative Fine-Tuning on high-quality data archive strong mathematical reasoning performance, like o1? Harvard invented rStar in August 2024, and now Microsoft gives us rStar Math for small LM. Technical tutorial available here youtu.be/s3xeXteLgzA?...

submitted 41 days ago • 0 comments

What is the latest tech in AI reasoning? Today a new research paper was released with code, where we implement autonomous Chain-of-Thoughts (CoT). For more complex causal reasoning. But there are limits. Have a look: youtu.be/FR8oE8chp7c?...

submitted 42 days ago • 0 comments

Tiny Language Models as the core of intelligent AI edge devices? Watch when scaling breaks down and we have to be creative with our training of real tiny LMs. youtu.be/TU19Orwu4jE?...

submitted 43 days ago • 0 comments

What do you prefer? In-context Learning with LCLM or classical Graph-RAG? Latest insight and comparison: youtu.be/tHrCE0gjq3I?...

submitted 43 days ago • 0 comments

Multi Agent AI systems are real powerhouses. The simplest way to code multiple AI agents? I experienced today "smolagents", and integrate multiple agents in a dynamic schema. Smolagents are simple to code, you can add tools and they are integrated in Python ecosystem youtu.be/oqqQ64BRyJE?...

submitted 49 days ago • 0 comments

2025, the year of AI agents?! You want to learn how to apply a modern version of AI agents? Huggingface just released their "smolagents". Easy to learn - smolagents empower modern multi AI Agent systems & you can build your own tools for complex tasks! Latest tech. youtu.be/tBXAwCH6rcU?...

submitted 50 days ago • 0 comments

META is flooding insta with AI Influencer? AI bots on all social platforms? Is there a deeper reason? Like financial incentives? My 🙂 simple thoughts ...

submitted 53 days ago • 0 comments

For 4 years RAG was the go-to solution to integrate external knowledge to our LLMs. Now we have a new, better solution: Cache Augmented Generation CAG. No more RAG. No vector stores, no cosine sim, just pure flash attention. youtu.be/NaEf_uiFX6o?...

submitted 54 days ago • 0 comments

Ever thought of an LLM that continuously learns? No additional fine-tuning, no new adapters, no reindexing of the vector store with new data added? Intel, Salesforce and Univ of Illinois developed a new ICL based method for continual learning. Great ideas, with no RAG youtu.be/qPxAlwkDguo?...

submitted 55 days ago • 0 comments

The new inference compute, like the o3 model, has significant impact on the fine-tuning and RL method. Test time compute needs to be aligned to SFT and RL. A new SFT and RL method for max inference reasoning. Watch here youtu.be/0xe0xal4pOY?...

submitted 55 days ago • 0 comments

New Knowledge Graph based RAG! Short tutorial on 4 classical KG- based RAG systems and the latest /best one: SimGRAG. It outperforms the classical KG-based RAGs and is a simple 2 step implementation. Code and explanation of new SimGRAG: youtu.be/aPsfAkrkma0?...

submitted 58 days ago • 0 comments

The inference reasoning performance of o3 is amazing, after 13 min. Given the latest tech report by OpenAI on DA, I set out to build a similar reasoning capability for my 7B model. First step: Create the SFT and RL optimized training data sets for Inference CoT reasoning. youtu.be/oUj4gYsNtdA?...

submitted 60 days ago • 0 comments

I explored the dependencies of SFT and RL in relation to inference reasoning of the new o3 model by OpenAI. Turns out, for the best performance you could implement a coherent training methodology from SFT to RL to Test Time Reasonings of #o3. 🧪 youtu.be/def4p1SEO3A?...

submitted 60 days ago • 0 comments

o3 is primarily a lighthouse signal to investors, that AI is not saturating, AI is not hitting a wall, but showing global investors new profitable paths to the future. And global Investors will continue to invest. It is a mind game - with trillions of dollars, but primarily on the US market.

submitted 63 days ago • 0 comments

Detailed test data for the new o3 model by OpenAI indicate the importance of test time compute, with 13 minutes for o3 high performance. If pre-training LLMs hits a wall, inference compute might be the gate to the future of AI. youtu.be/YjyLBabHQiQ?...

submitted 63 days ago • 0 comments

Just look at the performance data of openAI's new o3. Although costs from $20 up to thousands of $ per single query seem oriented towards companies (pharma, biotech, finance, ..).

submitted 63 days ago • 0 comments

ChatGPT sys are the decoder-only part of a transformer. The missing part is called BERT, the encoder-only part of a transformer. Today, after 6 years, we got an update: ModernBERT. BERT is the main transformer to build RAG systems, so you better update. Tech video here: youtu.be/Z1Dl3juwtSU?...

submitted 64 days ago • 0 comments

A Large Language Model without a Tokenizer? Is this possible in #AI? Welcome to an explanatory deep dive into the layer calculations of a Byte Latent Transformer - BLT. Unique? With an entropy calculation of the new byte which determines latent batch boundaries. youtu.be/KZfGgmtQFh0?...

submitted 64 days ago • 0 comments

You know that Google's EVO 2 is outperforming SORA as a text 2 video generator. But now we can build "world models" with #AI, where when we turn virtually into an unknown street, AI will generate a #video with a continuation of the 3dim world. John Hopkins shows the way: youtu.be/Mkn3KYYv5mk?...

submitted 65 days ago • 0 comments

If you want to analyze videos, find a particular object in the video or a special action, now we have a free option to do so. Gemini 2.0. a brand new video analyzer, where your prompt defines the action to search for in the video (visual content). #AI #Video 🧪 youtu.be/FY_Mqnc_6X8?...

submitted 65 days ago • 0 comments

When I read, that META developed a new Video #AI model, I was excited. When I read, that #META cooperated with #Stanford Univ for a Video AI that was just a 3B or a 7B, it was exciting. But the technical paper uncovered a surprise. youtu.be/OHy5VithNKc?...

submitted 67 days ago • 0 comments

Wow! I just listened to my YouTube video in the German version, that Google automatically dubbed. It is horrible. Smile. Please click on "original soundtrack", because it switches automatically to your regional language, if you play it here. Translation in no way understands technical terms.

submitted 68 days ago • 0 comments

What if you have to solve a specific task, like optimize organic solar cell molecules? General AI models fail. GNN fails. Now we design task specific AI NN architecture, combine transformer w GNN, build fusion layers and stack those unique layers. youtu.be/5e18nTD1aP8?...

submitted 69 days ago • 0 comments

Forget classical #AI w/ LLMs. Forget next token prediction with autoregressive models. #META experiments with diffusion based Large Concept Models, where Transformer build new concept embeddings as vectors and optimize denoising in Diffusion LCM. 🧪 youtu.be/y1MG0BCf3UU?...

submitted 70 days ago • 0 comments

Test-time compute of OpenAI's o1 model is impressive? Now we have a new reasoning optimization for inference (TTT): Forest-of-Thoughts. Because Chain-of-Thoughts CoT is not powerful enough: youtu.be/4bslHbUH5xc?...

submitted 71 days ago • 0 comments

Adaptive Intelligence for small LM is the NextGen challenge for AI 2025. Especially when to optimize adaptive ICL and when to apply Test-Time training adaptation (see video). Shifted data distribution are the key parameter. Detailed info here youtu.be/zmuNud2oBbE?...

submitted 72 days ago • 0 comments

My YouTube videos are now automatically dubbed from English to German, French, Italian, Spanish, and Portuguese. Thank you YouTube. You can experience it live m.youtube.com/@code4ai/vid...

submitted 72 days ago • 0 comments

You do not want to pay for fine-tuning of your LLM? New research into in- context learning ICL shows, there is a sweet spot for LLMs, regarding high ICL performance. But what LLMs have the best ICL performance? Online Test in my community tab. More info here youtu.be/f_z-dAQb3vw?...

submitted 73 days ago • 0 comments

Researcher from Princeton Univ found a phase transition in LLMs. A phase transition regarding the learning performance of Large Language Models. So no more expensive fine-tuning for learning new knowledge? It is a delicate balance with newly discovered scaling laws. youtu.be/f_z-dAQb3vw?...

submitted 74 days ago • 0 comments

The brand-new Llama 3.3? Llama 3.3 70B is just an old 3.1 70B finetuned on a new dataset, with a cutoff date of Dec 2023. Easy to do, cheap to generate and news media report on a breakthrough. The world is just looking for hype, not for content.

submitted 75 days ago • 0 comments

We train our Ai systems on millions of examples. And hope the machine will discover the hidden patterns. But what happens when we train LLMs directly on higher-level abstract reasoning structures? More ideas on this in my new video youtu.be/bSLc-dhn9vg?...

submitted 76 days ago • 0 comments

Small Language Models need smarter learning procedures. Today I explore in-context learning ICL for improved #AI reasoning capabilities of SLM. Cool AI research papers on ICL also in my video youtu.be/G_Z3dhsCbgM?...

submitted 76 days ago • 0 comments

If you ask yourself: what is this new #reinforcement fine-tuning for your #o1 #AI model, that #OpenAI reveales on day 2? My explanatory video on optimizing the fine-tuning phase with advanced non-distribution constrained RL methods you find here: youtu.be/dr5AL1abQW0?...

submitted 77 days ago • 0 comments

Google uncovers two new AI models. Genie 2 is an autoregressive latent diffusion model to generate 3dim game environments. # AI #gamedev PaliGemma 2 is the latest update of it's Vision - Language Model. All technical details and python code (plus online demo) is here: youtu.be/SQRu0cqmWqE?...

submitted 78 days ago • 0 comments

Amazon now has new AI models, dev by #Amazon. They claim, that NOVA Pro excels at analyzing financial reports, and I took the new #AI systems to a test. The result was not what I expected: 🧪 youtu.be/aMdbkzdoMrI?...

submitted 78 days ago • 0 comments