Wiredframe Radar logo

Category intelligence

Social Media Briefing — April 6, 2026

280 current items analyzed and ranked.

Executive synthesis

Social Media Summary

A fundamental debate about AI architecture headlined the day. Yann LeCun challenged Elon Musk and xAI's language-based reasoning approach, arguing real thinking manipulates mental models in continuous space — not tokens. François Chollet reinforced this from a different angle, arguing science achieves extreme generalization through symbolic compression, not brute-force language processing.

  • Stanford's MIRAGE paper sent shockwaves: major vision models including GPT-5, Gemini, and Claude score 70-80% on benchmarks even with images removed, exposing deep evaluation flaws
  • Andrej Karpathy publicly criticized X API pricing ($200 for 30 min of hacking) and poor documentation for AI agents, sparking discussion about platform readiness for agentic use cases
  • Greg Brockman announced Codex app server for building custom agentic apps, while Ethan Mollick highlighted that inference-time scaling hasn't plateaued — raising token limits from 3M to 10M tripled Codex autonomous work capacity

The on-device AI space saw intense interest around Google's Gemma 4, with Mollick calling it 'GPT-4ish quality' but warning small models lack the judgment needed for real agentic workflows. Meanwhile, GLM-OCR, a tiny 0.9B Chinese model, topped document reading benchmarks against models 100x its size.

Key Themes

World Models vs. Language-Based Reasoning · 5AI Benchmarking & Evaluation Flaws · 1Agentic AI Infrastructure · 4Scaling Laws and Inference-Time Compute · 2On-Device AI and Small Model Limitations · 5AI Platform and API Economics · 2Claude Code & Knowledge Management Workflows · 3Gemma 4 Community Reception · 4Inference-Time Compute / Token Scaling · 2Gemma 4 & On-Device AI · 3

Primary evidence

Top Ranked Signals

Social Twitter Apr 5

@elonmusk Thinking in language has limited applications, largely in coding and mathematics where the...

By @ylecun

92 score
AI Analysis

LeCun responds to Elon Musk arguing that thinking in language has limited applications (mainly coding/math), and that real thinking manipulates mental models in abstract continuous representation space. Asks if xAI will adopt JEPA.

@elonmusk Thinking in language has limited applications, largely in coding and mathematics where the language itself can help reasoning. But, as I've been saying for years, thinking manipulates mental models in abstract (continuous) representation space. Soooo, xAI gonna use JEPA now?
JEPAworld modelslanguage vs thoughtreasoningxAILeCun vs Musk
Social Twitter Apr 5

@chrisparkX I think it's a good direction (for Read endpoints, not for Write), I tried to use it for...

By @karpathy

88 score
AI Analysis

Karpathy criticizes X API's excessive pricing ($200 for 30 min of hacking), poor documentation for AI agents, and notes XMCP is not mentioned in docs despite being announced.

@chrisparkX I think it's a good direction (for Read endpoints, not for Write), I tried to use it for a project ~2 weeks ago but about 30 minutes of hacking around cost me $200, the pricing is imo really excessive. The docs were hard to ingest into agents because it's a lot of individual short pages, I think a big intro markdown doc, or a few of them behind simple curl locations. Also, the current version of docs seems to have no mention of XMCP? Or at least the Search / Grok Assistant seems to s
X APIAI agentsdeveloper experienceXMCPplatform economics
Social Twitter Apr 5

Stanford just proved the biggest AI vision models are actually blind. The paper is called MIRAGE. ...

By @AlphaSignalAI

88 score
AI Analysis

Stanford's MIRAGE paper reveals that major AI vision models (GPT-5, Gemini, Claude) score 70-80% on benchmarks even when all images are removed. Models fabricate detailed descriptions of non-existent images. A tiny 3B text-only model beat frontier models and radiologists, exposing that 74-77% of 'vision' questions don't actually test vision.

Stanford just proved the biggest AI vision models are actually blind. The paper is called MIRAGE. They removed every image from 6 major benchmarks. GPT-5, Gemini, and Claude still scored 70-80% accuracy. The models never noticed the images were gone. They kept describing photos in detail. Diagnosing conditions from X-rays that weren't there. Stanford calls it the "mirage effect." The models built a fake reality and reasoned from it confidently. To prove it, they trained a tiny 3B t
multimodal AI evaluationAI benchmarking flawsAI safetyhallucinationmedical AI
Social Twitter Apr 5

Science went from the initial observation of radioactivity to a working atom bomb over 47 years via ...

By @fchollet

85 score
AI Analysis

Chollet argues that science achieves extreme generalization through symbolic compression — citing going from radioactivity to the atom bomb via ~9 key experiments and concise symbolic models, contrasting this with data-driven approaches.

Science went from the initial observation of radioactivity to a working atom bomb over 47 years via only about 9 distinct key experiments -- extremely few data points -- and symbolic models concise enough they would fit on a single page. This is what extreme generalization looks like, and it powered entirely by symbolic compression. Turn a handful of data points (deliberately collected) into a tractable plan to completely reshape reality, by reverse-engineering the causal symbolic rules behind
symbolic reasoninggeneralizationscientific discoveryAI philosophydata efficiency
Social Twitter Apr 5

Unappreciated fact is the second scaling law does not seem to completely plateau in many tasks: thro...

By @emollick

82 score
AI Analysis

Emollick argues the 'second scaling law' (inference-time compute) doesn't plateau for many tasks — more tokens to reasoning models yields better answers, and benchmarks are limited by token usage.

Unappreciated fact is the second scaling law does not seem to completely plateau in many tasks: throw more tokens at a reasoning AI model and get better answers, especially with a simple harness. Benchmark performance is actually limited by token usage. t.co/jP3TgIsT2T t.co/DQs6OJvNTy
scaling lawsinference-time computereasoning modelsbenchmarks
Social Twitter Apr 5

Gemma 4 E4B is impressive for an on-device LLM. GPT-4ish quality, and expect hallucinations. Here i...

By @emollick

80 score
AI Analysis

Following yesterday's News coverage of the Gemma 4 launch, Emollick demos Gemma 4 E4B on-device, calling it 'GPT-4ish quality' and showing real-time text generation including rhyming verse about sociological theories.

Gemma 4 E4B is impressive for an on-device LLM. GPT-4ish quality, and expect hallucinations. Here is: “List five sociological theories starting with u and what they are. Then describe them in a rhyming verse” Its in real time, the last is a little bit of a stretch, but not bad! t.co/nJ5HuQsFSB
Gemma 4on-device AImodel evaluationopen-source AI
Social Twitter Apr 5

I am impressed by Gemma 4, there’s a lot of power for an on-device model at fast speeds. But I am no...

By @emollick

78 score
AI Analysis

Following yesterday's News coverage of the Gemma 4 launch, Emollick praises Gemma 4's speed and power as an on-device model but argues small models lack the judgment, self-correction, and accuracy needed for real agentic workflows.

I am impressed by Gemma 4, there’s a lot of power for an on-device model at fast speeds. But I am not convinced you can get real agentic workflows out of a small model on device. So much depends on model judgement, self-correction, and accuracy. Small models are too weak there. t.co/Adi9vlYL24
Gemma 4on-device AIagentic workflowsmodel capability limits
Social Twitter Apr 5

A peanut-sized Chinese model just dethroned Gemini at reading documents. GLM-OCR is a 0.9B paramete...

By @AlphaSignalAI

78 score
AI Analysis

GLM-OCR, a 0.9B parameter Chinese vision-language model, achieves #1 on OmniDocBench V1.5 (94.62 score), outperforming models 100x its size. Open-source, handles complex tables, handwriting, math formulas. Runs locally on edge devices via Ollama using multi-token prediction.

A peanut-sized Chinese model just dethroned Gemini at reading documents. GLM-OCR is a 0.9B parameter vision-language model. It scores 94.62 on OmniDocBench V1.5, ranking #1 overall. For context, it outperforms models 100x its size. 100% open-source. It works in two stages. 1. A layout engine detects every region in a document. 2. Each region gets read in parallel. The model predicts multiple tokens per step instead of one. That's what makes it so fast at small size. It handles things
small modelsOCRopen-source AIedge AIefficiency
Social Twitter Apr 5

I'm a knowledge base MONSTER in Claude Code right now. Introducing: Claudeopedia. 1) I took @kar...

By @alliekmiller

78 score
AI Analysis

Building on Karpathy's viral Social post about LLM knowledge bases, Allie K Miller introduces 'Claudeopedia' - a knowledge base system built in Claude Code combining Karpathy's llm-wiki idea with custom skills for knowledge management, visualization, and automated assumption-checking in Obsidian.

I'm a knowledge base MONSTER in Claude Code right now. Introducing: Claudeopedia. 1) I took @karpathy's 'llm-wiki' idea doc (90% of this, so the biggest credit goes to @karpathy) and 2) Combined it with the /last30days skill (HT @mvanhorn) and 3) Added a /wiki skill with screenshot and download arguments to transfer raw inputs faster and 4) Built an interactive visualization to search my knowledge base (with date ranges to compare knowledge over time!) 5) Set up a "question your assumption
Claude Codeknowledge managementAI workflowsKarpathy llm-wikienterprise AIObsidianpersonal productivity
Social Twitter Apr 5

Real world example: raising token limits from 3M to 10M tripled the amount of work that Codex could ...

By @emollick

75 score
AI Analysis

Emollick shares that raising token limits from 3M to 10M tripled Codex's independent work on cybersecurity tasks (3.1 to 10.5 hours).

Real world example: raising token limits from 3M to 10M tripled the amount of work that Codex could do independently on cybersecurity tasks. From 3.1 hours to 10.5 hours.
token scalingagentic codingOpenAI Codexcybersecurity
Social Twitter Apr 5

@vitrupo Language is a way to communicate thoughts. What good is language if you don't have thoughts...

By @ylecun

72 score
AI Analysis

LeCun argues that language is merely communication of thoughts, and thoughts are manipulations of mental models — implying language alone is insufficient for intelligence.

@vitrupo Language is a way to communicate thoughts. What good is language if you don't have thoughts? And what are thoughts if not manipulations of mental models of the world?
world modelslanguage vs thoughtAI philosophyJEPA