Claude 4: Full 120 Page Breakdown … Is it the Best New Model?
Not only did I get early access and ran my own tests, as per the title I read both the 120 page Claude 4 Opus and Claude 4 Sonnet System Card, and 25 page report on ASL-3 being triggered, plus the 2 hour launch video, and surrounding coverage. Ft. coding tests, Simple, twitter controversies, deep alignment coverage, spiritual bliss and much more!https://80000hours.org/aiexplainedChapters: 00:00 - Introduction01:12 - 3 Quick Controversies02:42 - Benchmark Results 04:20 - 120 page Card 20 Highlights10:07 - Coding Test11:27 - Model Welfare and Spiritual Bliss13:29 - ASL-3Claude Card: https://www-cdn.anthropic.com/4263b940cabb546aa0e3283f35b686f4f3b2ff47.pdf?s=09ASL 3:https://www-cdn.anthropic.com/807c59454757214bfd37592d6e048079cd7a7728.pdfTweets: https://x.com/fish_kyle3/status/1925597284546629753https://x.com/EMostaque/status/1925624164527874452?ref_src=twsrc%5Egoogle%7Ctwcamp%5Eserp%7Ctwgr%5EtweetCursor Says State of the Art for Coding: https://x.com/cursor_ai/status/1925594428095561941Benchmarks: https://www.anthropic.com/news/claude-4
--------
19:04
Google Takes No Prisoners Amid Torrent of AI Announcements
Google just announced at least 12 things that are each worthy of a video, but here are the top I/O highlights. From Veo 3 to Deep Research now being useable, Deep Think breaking records to Gemini Diffusion, Gemini 2.5 Flash changing how AI is priced and GemmaVerse, SynthID Detector and Imagen 4. And even this intro is missing other announcements covered in the vid! And yes, they’ll be plenty of Veo 3 clips to enjoy…https://80000hours.org/aiexplainedAI Insiders ($9!): https://www.patreon.com/AIExplainedChapters:00:00 - Introduction00:48 - Veo 302:10 - Gemini 2.5 Flash03:13 - Universal Assistant03:47 - Usage Skyrockets + OpenAI dig04:51 - Gemini Pro Deep Think06:21 - Overviews and AI Mode07:26 - Deep Research Updates (new) + Jules 08:53 - Make and Deploy Apps with Gemini09:12 - Imagen 4 10:00 - Gemini Diffusion11:46 - Try It On12:17 - SynthID Detector13:30 - GemmaVerse, SignGemma, Gemma3n, medGemma14:24 - Outro + ClipsEvent: https://www.youtube.com/watch?v=o8NiE3XMPrMNtaive Audio: https://aistudio.google.com/generate-speechGemini Diffusion: https://deepmind.google/models/gemini-diffusion/#capabilities New Gemini 2.5 Flash: https://deepmind.google/models/gemini/flash/SignGemma (See end of this vid): https://www.youtube.com/watch?v=GjvgtwSOCaoDeep Think: https://blog.google/technology/google-deepmind/google-gemini-updates-io-2025/#flash-improvementsGoogle Parallel Sampling: https://www.patreon.com/posts/next-level-good-127441188Price Plans: https://blog.google/products/google-one/google-ai-ultra/Imagen 4 Benchmarks: https://deepmind.google/models/imagen/Jules: https://jules.google/SynthID Detector: https://blog.google/technology/ai/google-synthid-ai-content-detector/Veo 3 Benchmarks: https://deepmind.google/models/veo/evals/MedGemma: https://deepmind.google/models/gemma/medgemma/Build Apps: https://aistudio.google.com/appsNon-hype Newsletter: https://signaltonoise.beehiiv.com/
--------
17:07
AI Improves at Self-improving
AlphaEvolve is not the first system to exhibit self-improvement, but it may be the most impressive yet. AI is literally improving the hardware, architectures, data and training methods of AI itself. A deep dive into the paper, drawing on two previous interviews and 5 other papers. Plus a snippet on OpenAI’s new Codex system.Gray Swan: http://app.grayswan.ai/ai-explainedAI Insiders ($9!): https://www.patreon.com/AIExplainedChapters:00:00 - Introduction00:27 - AlphaEvolve05:23 - Limitation06:10 - Achievements08:21 - Future Improvements13:30 - Quirks16:34 - Final ThoughtsAlphaEvolve release: https://deepmind.google/discover/blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/Paper: https://storage.googleapis.com/deepmind-media/DeepMind.com/Blog/alphaevolve-a-gemini-powered-coding-agent-for-designing-advanced-algorithms/AlphaEvolve.pdfTerence Tao Quote: https://mathstodon.xyz/@tao/114508029896631083Nature Article: https://www.nature.com/articles/s41586-022-05172-4MIT Article: https://www.technologyreview.com/2025/05/14/1116438/google-deepminds-new-ai-uses-large-language-models-to-crack-real-world-problems/AI Co-Scientist: https://arxiv.org/pdf/2502.18864OpenAI Codex: https://openai.com/index/introducing-codex/70% of Pull Requests: https://x.com/slow_developer/status/1920920456393028027Amodei Essay: https://www.darioamodei.com/essay/machines-of-loving-graceOpenAI Jason Wei Tweet: https://x.com/_jasonwei/status/1923091260354531612PromptBreeder: https://arxiv.org/pdf/2309.16797DrEureka: https://arxiv.org/pdf/2406.01967FT DeepMind: https://www.ft.com/content/4e497a91-670a-4f69-be4a-18e247daba3eNon-hype Newsletter: https://signaltonoise.beehiiv.com/
--------
17:41
o3 breaks (some) records, but AI becomes pay-to-win
A green card, o3 vs Gemini 2.5, 6 Benchmarks and a whole bunch of my thoughts on what on earth is happening in AI, from here to 2030. Plus, how AI is becoming pay-to-win, and why. Crazy times, 14 mins probably wasn’t enough.https://app.grayswan.ai/ai-explainedAI Insiders ($9!): https://www.patreon.com/AIExplainedChapters:00:00 - Introduction00:33 - FictionLiveBench01:37 - PHYBench02:14 - SimpleBench02:54 - Virology Capabilities Test03:13 - Mathematics Performance04:29 - Vision Benchmarks05:43 - V* and how o3 works06:44 - Revenue and costs for you08:54 - Expensive RL and trade-offs 09:40 - How to spend the OOMs13:27 - Gray Swan ArenaGreen Card: https://techcrunch.com/2025/04/25/an-openai-researcher-who-worked-on-gpt-4-5-had-their-green-card-denied/PHYBench: https://arxiv.org/pdf/2504.16074Virologytest: https://www.virologytest.ai/How o3 Vision Works: https://arxiv.org/pdf/2312.14135 https://x.com/sainingxie/status/1912570624523829573Visual puzzles: https://neulab.github.io/VisualPuzzles/Fiction Bench: https://x.com/ficlive/status/1912863028141244850https://geobench.org/https://simple-bench.com/AIME 2025: https://openai.com/index/introducing-o3-and-o4-mini/USAMO: https://x.com/mbalunovic/status/1914398518896193747NaturalBench: https://linzhiqiu.github.io/papers/naturalbench/Where’s Waldo: https://uk.pinterest.com/pin/492792384225896298/IMO and AlphaProof:https://deepmind.google/discover/blog/ai-solves-imo-problems-at-silver-medal-level/Crazy Revenue: https://www.theinformation.com/articles/openai-forecasts-revenue-topping-125-billion-2029-agents-new-products-gain?rc=sy0ihqNumber of Users: https://www.theinformation.com/briefings/googles-gemini-user-numbers-revealed-court?rc=sy0ihqSubscriptions pay to win: https://www.forbes.com/sites/paulmonckton/2025/04/23/google-leak-reveals-new-gemini-ai-subscription-levels/GPU Trade-offs: https://x.com/sama/status/1915098951067554030RL Scale-up Amodei: https://www.darioamodei.com/post/on-deepseek-and-export-controlsLog-linear Returns: https://x.com/bobmcgrewai/status/18952282919819432652030 Scaling: https://epoch.ai/blog/can-ai-scaling-continue-through-2030Model Size: https://x.com/slow_developer/status/1874554473256997201Adam on AGI: https://x.com/TheRealAdamG/status/1913998366632968381Papers on Patreon: https://arxiv.org/pdf/2502.01839https://arxiv.org/pdf/2504.13837Chollet Quote: https://x.com/fchollet/status/1912934762580447447OpenSim: https://opensim.stanford.edu/Non-hype Newsletter: https://signaltonoise.beehiiv.com/
--------
14:33
o3 and o4-mini - they’re great, but easy to over-hype
Critical analysis of the two most powerful new models behind ChatGPT, o3 and o4-mini. Not just the system cards, benchmarks, and my own tests, but some you may not have seen before. Yes, they can whip up amazing front-end in a few seconds, but you always have to ask what is in their data. Either way, they prove the gains from RL are just beginning…https://weave-docs.wandb.ai/?utm_source=sponsorship&utm_medium=simple_bench&utm_campaign=ai_explainedAI Insiders ($9!): https://www.patreon.com/AIExplainedChapters:00:00 - o3 and o4-minihttps://simple-bench.com/Plus, Teams and Pro, plus token count: https://x.com/btibor91/status/1912568994512662679System Card: https://openai.com/index/o3-o4-mini-system-card/Release Notes: https://openai.com/index/introducing-o3-and-o4-mini/https://deepmind.google/technologies/gemini/pro/https://x.com/DeryaTR_/status/1912558350794961168https://x.com/polynoamial/status/1912564068168450396API Pricing:https://openai.com/api/pricing/https://aider.chat/docs/leaderboards/Non-hype Newsletter: https://signaltonoise.beehiiv.com/
Covering the biggest news of the century - the arrival of smarter-than-human AI. From the author of Simple Bench, which reveals the remaining gap between LLM and human reasoning. Hype-free, and the British accent is a freebie bonus.
Brasil