Evalds Urtans

Podcast with @AndrisKulbergs - Prime Minister candidate - Can AI replace politicians? Full episode: youtu.be/DHwvgHEuMnE

@elnormous :D

Zvanu krāpniecības 3. sērija. Šajā sērijā krāpniece ar mani sazinās WhatsApp, bet es tēloju pilnīgu idiotu, kas viņu nedaudz nokaitina.

Fully AI agentic cafe = chaos + over budget 😄 "Mona's inventory missteps have become a running joke among the staff. The agent ordered 6,000 napkins, four first-aid kits, 3,000 rubber gloves, and canned tomatoes that no dish on the menu calls for. The baristas set up a "Hall of Shame" — a shelf visible to customers displaying Mona's most bewildering purchases, including 9 liters of coconut milk and industrial-sized trash bags" perplexity.ai/discover/you/a…

Sweden’s approach to AI appears impressive—at least until you learn what’s really happening behind the scenes. youtu.be/r-_wS_adkoc

Sweden’s AI Policy Seems Impressive Until You Hear the Inside Story. Oscar Chröisty on our 60th episode of @Ai_Ranch podcast. Youtube - youtu.be/r-_wS_adkoc Spotify - open.spotify.com/episode/1hEeF3…

AI news - April 2026 🚀🚀 youtube.com/watch?v=6vZg2i…

Psyhology & AI 🔮! Gustavo Moreno e Mello on AI Ranch from Oslo, Norway 🇳🇴! youtube.com/watch?v=6TjlVv…

@karpathy True, how about endless clutter of information inside OpenClaw soul.md identity.md, would’nt ir work much better with iterative and short instructions instead of trying to force unreasonable amount of instructions?

One common issue with personalization in all LLMs is how distracting memory seems to be for the models. A single question from 2 months ago about some topic can keep coming up as some kind of a deep interest of mine with undue mentions in perpetuity. Some kind of trying too hard.

@jahirsheikh8 AI in Europe - AI Ranch 🤠 @airanch?si=3o6uscAxmpDsltgK" target="_blank" rel="nofollow noopener">youtube.com/@airanch?si=3o…

Best YouTube Channels To Learn AI in 2026 (No BS) 1. Fundamentals – 3Blue1Brown 2. Deep Learning – Andrej Karpathy 3. AI Research – Yannic Kilcher 4. Practical AI – AssemblyAI 5. LLMs – AI Explained 6. ML Theory – StatQuest 7. Papers Simplified – Two Minute Papers 8. GenAI – Matthew Berman 9. AI Agents – Nicholas Renotte 10. Applied ML – Krish Naik 11. PyTorch – Aladdin Persson 12. Math for ML – Serrano Academy 13. Industry Insights – Lex Fridman 14. Real-world AI – DeepLearningAI

Cik ātri attīstās AI inteliģence? Es sekoju līdzi Humanity's Last Exam (HLE), kur 1000 zinātnieki no 50 valstīm salikuši kopā 2500 jautājumus, lai novērtētu AI "intelektu". agi.safe.ai Šeit ir grafiks par pēdējiem 2 gadiem. No <10% līdz nu jau 45.9%.

What an amazing story! This is how AI can make a difference! It also shows how terribly slow our current medical research & approval process is at bringing life-saving treatments to patients. With AI we can accelerate all this by orders of magnitude & save so many lives!

Recent research from Harvard and Goodfire AI reveals that LLMs don't actually need Chain of Thought reasoning for simple tasks. Although step-by-step thinking technically works, the models are essentially playing a Theatre — they're already capable of producing correct responses more quickly using fewer tokens. However, when it comes to harder problems, there are real breakthrough moments (aha moments) during the reasoning process that genuinely enhance the quality of outputs. Notably, models with more parameters reach these breakthroughs earlier.

We are only just entering the AI era. Below is a framework for identifying untapped business niches. Although modern AI models could, in principle, take over much of this work, most opportunities remain unexplored. Sections marked in red show what is already automated, while those in blue highlight what can yet be automated. Report: anthropic.com/research/labor…

Excited to introduce the asya.ai open-source project AsyaChatUI 🎉🎉🎉 to deploy an LLM chat UX like ChatGPT on your own servers, using Azure Private GPT, Google Gemini API, Anthropic API, etc. This solves multiple major issues for enterprises: 1. Host everything on your servers, with no need for external providers. Control access and user rights. 2. Do not pay per user—just for what you have been using—massive savings. 3. Use your own Microsoft tenant and agreement to support GDPR compliance, eliminating the need to delete sensitive data from prompts and files used in chat. There are many more benefits, deploy, start using & contribute! github.com/asya-ai/asya-c…

@didzisdejus 💯😃

Šis ir izcils konts. Iesaku!

ChatGPT tip: simply adding the phrase ‘Look at examples in the “Examples” section’ can raise an LLM’s accuracy by up to 44.62%, even when no examples are actually added :D. A new scientific study argues that hallucination in LLMs may not just be a flaw—it could be a still-mysterious feature. The researchers dub this phenomenon ‘Null-Shot Prompting,’ and report that it works especially well on chat-tuned Transformer models tackling math, reading-comprehension tasks, and even hallucination-detection problems. aclanthology.org/2024.emnlp-mai…

Yngvar Ugland on @Ai_Ranch - Do you trust your bank? Chek out the episode on youtube - youtu.be/ZH2C3XtfuCc Or spotify - open.spotify.com/episode/1pRuHm… #AI #Banks #DNB #AIRanch #DigitalTrust

Tip for using RAG systems like NotebookLM, LLamaIndex, and others: The fewer documents you add, the better the semantic search will work, so do not add unnecessary clutter to NotebookLM or other RAG systems. Google DeepMind has shown that for an embedding space with dimension d, the total number of distinct top-k document sets that can be produced for any query is capped by that very dimension. Consequently, extra training, more data, or larger models cannot surpass this boundary—it represents a mathematical limit rather than an engineering challenge. Simply expanding the model’s dimensionality is not a straightforward remedy either. Owing to the curse of dimensionality, as the number of dimensions grows, the angles between vectors converge toward 90°, which effectively erases meaningful differences between documents. arxiv.org/pdf/2508.21038

Whenever possible, start a new chat or use the edit option rather than continuing an existing conversation. Studies show that while models respond correctly to single, isolated queries about 90 % of the time, their accuracy sinks to roughly 65 % during extended back-and-forth dialogue. Researchers refer to this drop-off as the “lost in conversation” effect—once an early mistake occurs, the model has difficulty recovering. arxiv.org/pdf/2505.06120