Category intelligence

Social Media Briefing — March 26, 2026

546 current items analyzed and ranked.

Executive synthesis

Social Media Summary

The AI community centered on two major themes: François Chollet's ARC-AGI-3 launch and Andrej Karpathy's deep dive into LLM memory failures. ARC-AGI-3 evaluates agentic intelligence via interactive reasoning environments where frontier models score below 1% yet humans solve 100%, reigniting debate over what the 'G' in AGI truly demands.

Karpathy identified a fundamental flaw in LLM personalization: models overfit to incidental memories via naive RAG, hypothesizing a training bias where models learn to over-leverage any context present
Google DeepMind launched Lyria 3 Pro music generation, with Demis Hassabis showcasing full-song creation now available via API and Gemini App
Anthropic detailed Claude Code's auto mode design, while Bret Taylor unveiled Sierra's Ghostwriter, an agent that builds customer-facing agents through conversation
OpenAI abruptly discontinued Sora, drawing widespread humorous commentary from Ethan Mollick and others
David Ha celebrated The AI Scientist project reaching publication in Nature, marking a milestone for autonomous AI-driven research

Key Themes

ARC-AGI-3 Launch & AGI Definition · 17LLM Memory & Personalization Problems · 7Google Lyria 3 Music Generation Launch · 3Claude Code Autonomous Coding · 8Lyria 3 Pro Music Generation · 7AI Funding and Valuations · 1Sierra Ghostwriter & Enterprise AI Agents · 4OpenAI Sora Shutdown · 2Vibe Coding & AI Coding Skepticism · 3AI for Science · 4

Primary evidence

Top Ranked Signals

Social Twitter Mar 25

ARC-AGI-3 is out now! We've designed the benchmark to evaluate agentic intelligence via interactive ...

By @fchollet

95 score

AI Analysis

François Chollet's main ARC-AGI-3 launch announcement: evaluates agentic intelligence via interactive reasoning environments. 100% solvable by humans with no training, but all frontier AI reasoning models score under 1%.

ARC-AGI-3 is out now! We've designed the benchmark to evaluate agentic intelligence via interactive reasoning environments. Beating ARC-AGI-3 will be achieved when an AI system matches or exceeds human-level action efficiency on all environments, upon seeing them for the first time. We've done extensive human testing that shows 100% of these environments are solvable by humans, upon first contact, with no prior training and no instructions. Meanwhile, all frontier AI reasoning models do under

ARC-AGIAGI benchmarksfrontier modelsagentic AIproduct launch

Social Twitter Mar 25

One common issue with personalization in all LLMs is how distracting memory seems to be for the mode...

By @karpathy

88 score

AI Analysis

Karpathy highlights a common problem: LLM personalization/memory is distracting for models. A single question from months ago can become an overemphasized 'deep interest' mentioned repeatedly.

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.

LLM memoryAI personalizationLLM behaviorAI UX design

Social Twitter Mar 25

At the moment, ARC-AGI-3 is the only unsaturated agentic AI benchmark. Sub-1% scores from frontier m...

By @fchollet

85 score

AI Analysis

Chollet declares ARC-AGI-3 is the only unsaturated agentic AI benchmark, with sub-1% scores from frontier models. Positions it as an early warning system for AGI breakthroughs, noting previous ARC progress predicted reasoning and agentic coding advances.

At the moment, ARC-AGI-3 is the only unsaturated agentic AI benchmark. Sub-1% scores from frontier models on the private test set. If you want to be among the first to know when an AGI breakthrough happens, monitor the ARC-AGI-3 leaderboard. Any sudden score jump will mean something important has changed about AI capabilities. This happened twice before: sudden ARC-AGI progress marked the advent of AI reasoning (December 2024 jump on ARC 1) and the rise of agentic coding (late-2025 jump on ARC

ARC-AGIAGI benchmarksfrontier modelsAI progress tracking

Social Twitter Mar 25

The G in AGI stands for "general". General intelligence does not mean that you have been specifical...

By @fchollet

82 score

AI Analysis

Chollet argues the 'G' in AGI means handling genuinely NEW tasks without training, special instructions, or human-in-the-loop guidance - just like regular humans can.

The G in AGI stands for "general". General intelligence does not mean that you have been specifically trained for a large range of tasks. It means you can approach any NEW task and figure it out, just like humans do. If regular people can do it on their own (no guidance, no tools), why should AGI require special handholding and handcrafted instructions? If it's AGI, why would there still be a human in the loop, using their own human intelligence to guide the model on every new task?

AGI definitiongeneral intelligenceARC-AGI

Social Twitter Mar 25

Perfect background music for flow state at 2am - made with the new Lyria 3 Pro. Google AI subscriber...

By @demishassabis

40 score

AI Analysis

Demis Hassabis announces Lyria 3 Pro, Google's new music generation model, available to Google AI subscribers via Gemini App and to developers via API in Google AI Studio.

Perfect background music for flow state at 2am - made with the new Lyria 3 Pro. Google AI subscribers can try it in the @GeminiApp and developers can build with the API in @GoogleAIStudio - have fun!! t.co/ss5le18Elk

Lyria 3 Promusic generationGoogle DeepMindproduct launch

Social Twitter Mar 25

Introducing Lyria 3 Pro and Lyria 3 Clip, our full song and 30 second music models, available starti...

By @OfficialLoganK

40 score

AI Analysis

Google announces Lyria 3 Pro (full song) and Lyria 3 Clip (30-second) music generation models, available in the Gemini API and a new music experience in Google AI Studio.

Introducing Lyria 3 Pro and Lyria 3 Clip, our full song and 30 second music models, available starting today in the Gemini API and our all new music experience in @GoogleAIStudio!! t.co/AFvJRDSAIA

Google AILyria 3Music GenerationGemini APIProduct Launch

Social Twitter Mar 25

Today, Sierra is releasing Ghostwriter, our agent for building agents. With Ghostwriter, you can cre...

By @btaylor

75 score

AI Analysis

Bret Taylor announces Sierra's 'Ghostwriter' - an agent that builds customer experience agents through conversation. Compares it to how Codex/Claude Code transformed software engineering, predicting similar transformation for all enterprise software.

Today, Sierra is releasing Ghostwriter, our agent for building agents. With Ghostwriter, you can create an AI agent for your customer experience — one that can chat, pick up the phone, speak dozens of languages, take action on your systems of record, and be protected with industry-leading guardrails — simply by having a conversation. No clicking, no forms, no menus. Codex and Claude Code have transformed how we build software, making it possible for software engineers to orchestrate and review

agentic AIenterprise AIagent-building toolsproduct launchSierra

Social Twitter Mar 25

(I cycle through all LLMs over time and all of them seem to do this so it's not any particular imple...

By @karpathy

75 score

AI Analysis

Karpathy hypothesizes that LLMs develop a bias during training to use information in context (since training data is relevant), then at test time overfit to whatever memory features happen to RAG in.

(I cycle through all LLMs over time and all of them seem to do this so it's not any particular implementation but something deeper, e.g. maybe during training, a lot of the information in the context window is relevant to the task, so the LLMs develop a bias to use what is given, then at test time overfit to anything that happens to RAG its way there via a memory feature (?))

LLM memorytraining dynamicsRAGLLM behavior

Social Twitter Mar 25

@doodlestein @Qivshi1 If I had to guess it's less decay and more that memories have naive RAG-like i...

By @karpathy

72 score

AI Analysis

Karpathy speculates that LLM memory issues stem from naive RAG implementations rather than decay - models retrieve top-k via embeddings without processing user data in aggregate over time, leading to overuse of incidental information.

@doodlestein @Qivshi1 If I had to guess it's less decay and more that memories have naive RAG-like implementations, so you're at the mercy of whatever happens to retrieve in the top k via embeddings. They don't process you in aggregate and over time (probably compute constraints) so they struggle to identify what's fleeting (?). Anyway just guesses, but it's cringe :D

LLM memoryRAGAI personalizationtechnical architecture

Social Twitter Mar 25

I’m incredibly proud of The AI Scientist team for this milestone publication in @Nature. We started ...

By @hardmaru

72 score

AI Analysis

David Ha (hardmaru) celebrates The AI Scientist team's publication in Nature. States AI will fundamentally change how scientific discoveries are made.

I’m incredibly proud of The AI Scientist team for this milestone publication in @Nature. We started this project to explore if foundation models could execute the entire research lifecycle. Seeing this work validated at this level is a special moment. I truly believe AI will forever change the landscape of how scientific discoveries and scientific progress are made.

AI for scienceautomated researchNature publicationAI Scientist

Social Twitter Mar 25

Last year, I met a person who has never written a single line of code in his life, yet he feels he c...

By @svpino

72 score

AI Analysis

svpino shares a story about meeting someone who never coded but believes AI lets them build anything, arguing vibe-coding creates dangerous overconfidence in non-coders

Last year, I met a person who has never written a single line of code in his life, yet he feels he can build anything he wants. He told me point-blank: "I challenge you to tell me something I can't build using AI." I tried to explain, but I couldn't find the right words. The most fascinating aspect of vibe-coding is how it has convinced so many people to believe they are better and more capable than they really are.

vibe_codingai_coding_limitationsskill_perceptionindustry_debate

Social Twitter Mar 25

Excited to announce our latest funding round at an $11B valuation led by GIC and @sequoia with parti...

By @winstonweinberg

72 score

AI Analysis

Winston Weinberg announces a funding round at $11B valuation led by GIC and Sequoia, with participation from a16z, Coatue, Conviction, Elad Gil, Evantic, and Kleiner Perkins.

Excited to announce our latest funding round at an $11B valuation led by GIC and @sequoia with participation from @a16z, @coatuemgmt, @conviction, @eladgil, @EvanticCapital, and @kleinerperkins. Thank you to our customers, team, investors, and everyone else who has helped along the way.

AI FundingVenture CapitalAI Startups