R[AI]DAR

Category intelligence

Social Media Briefing

All 518 current items, analyzed and ranked.

Executive synthesis

Social Media Summary

OpenAI Codex dominated the day's conversation. Sam Altman declared Codex is having a "ChatGPT moment" (697K views), while Greg Brockman claimed it has replaced his terminal as his primary computer interface after two decades — signaling a potential paradigm shift in developer tooling.

  • Anthropic published two notable research pieces: "introspection adapters" enabling LLMs to self-report learned behaviors and detect potential misalignment during training, and BioMysteryBench results showing Claude solving biology problems that stumped human expert panels
  • Gary Marcus went viral arguing Big Tech's collective AI spend exceeds the Manhattan Project monthly, calling it the "greatest capital misallocation in history" — he also shared a personal account of Altman lying under sworn testimony
  • swyx offered detailed technical praise of DeepSeek V4, highlighting its long-context efficiency innovations and confidence in skipping benchmark optimization
  • Ethan Mollick shared nuanced research showing AI-as-generic-study-tool hurts learning, but AI prompted as a tutor with teacher support yields large gains

Other notable signals: John Carmack discovered dramatic GPU performance cliffs at power-of-2 matrix boundaries, David Ha presented ICASSP 2026 research on "speak while thinking" voice AI, and Mayo Clinic's REDMOD AI detected pancreatic cancer 16 months before doctors on routine CT scans.

Key Themes

OpenAI Codex Launch Momentum · 8AI Safety and Alignment Research · 3AI Capital Allocation Bubble · 7AI for Science and Healthcare · 7DeepSeek V4 Analysis · 2GPU/CUDA Performance Optimization · 2AI in Education Research · 1Agentic AI & Agent Orchestration · 14Musk vs OpenAI/Altman Lawsuit · 15AI Agents & Inference Demand · 8

Primary evidence

All Ranked Signals

Showing 10 of 518 signals 2 remaining
Social Twitter Apr 29

Sheer insanity. Amazon, Google, Microsoft, and Meta collectively are spending more money than the Ma...

By @GaryMarcus

88 score
AI Analysis

Marcus argues Big Tech AI spending exceeds Manhattan Project monthly, calls it 'greatest capital misallocation in history' - no moats, no major profits, inevitable price war

Source details
Sheer insanity. Amazon, Google, Microsoft, and Meta collectively are spending more money than the Manhattan Project *every single month*. More than 12x the Manhattan Project every year. And what they have got to show for it? None are making major profits on AI; none has a technical moat; a massive price war is inevitable. And few of their customers are seeing major returns on investment. Greatest capital misallocation in history.
ai_bubblecapital_allocationno_moatindustry_dynamics
Social Twitter Apr 29

In new Anthropic Fellows research, we discuss “introspection adapters": a tool that allows language ...

By @AnthropicAI

42 score
AI Analysis

As covered in detail in Research yesterday, Anthropic announces 'introspection adapters' research — a tool allowing LLMs to self-report learned behaviors during training, including potential misalignment.

Source details
In new Anthropic Fellows research, we discuss “introspection adapters": a tool that allows language models to self-report behaviors they've learned during training—including potential misalignment.
AI safetyAI alignmentinterpretabilityAnthropic research
Social Twitter Apr 29

feels like codex is having a chatgpt moment

By @sama

82 score
AI Analysis

Continuing the Codex buzz from Social yesterday, Sam Altman says 'Codex is having a ChatGPT moment' - comparing the product's reception to ChatGPT's breakout

Source details
feels like codex is having a chatgpt moment
openai_codexproduct_launchai_coding
Social Twitter Apr 29

New on the Science Blog: We gave Claude 99 problems analyzing real biological data and compared its ...

By @AnthropicAI

82 score
AI Analysis

Anthropic tested Claude on 99 biological data analysis problems against expert panel. On 23 problems experts couldn't solve, Claude's latest models solved ~30% of those plus most of the rest.

Source details
New on the Science Blog: We gave Claude 99 problems analyzing real biological data and compared its performance against an expert panel. On 23 problems, the experts were stumped. Our most recent models solved roughly 30% of those—and most of the rest. t.co/BYqr76zxhk
AI for scienceAI benchmarksAnthropic researchAI surpassing human experts
Social Twitter Apr 29

Yes, just having students “use AI to study” hurts learning (a helpful assistant is not a tutor), but...

By @emollick

80 score
AI Analysis

Ethan Mollick shares research showing generic AI study help hurts learning, but AI prompted as a tutor with teacher support shows large positive effects in randomized trials

Source details
Yes, just having students “use AI to study” hurts learning (a helpful assistant is not a tutor), but using AI prompted to act like a tutor, especially with teacher support, seems to have large positive effects on learning in randomized trials. t.co/0HtjGC8eU0 t.co/tjkygnFCGc
ai_educationresearchai_implementation
Social Twitter Apr 29

IMO DeepSeek v4 demonstrated utter confidence and competence by not benchmaxxing, not focusing on so...

By @swyx

80 score
AI Analysis

Following yesterday's News coverage of DeepSeek V4, swyx gives strong praise to DeepSeek V4, highlighting their confidence in not benchmaxxing, demonstrating SOTA long context efficiency (CSA, HCA, mHC), pricing (Flash at 8% cost of Pro, which is 14% cost of Opus), and best open base models. Notes 'BYO posttraining' approach.

Source details
IMO DeepSeek v4 demonstrated utter confidence and competence by not benchmaxxing, not focusing on some BS final run cost, not even spending inference-optimal compute. just showed up, demonstrated SOTA long context efficiency techniques (CSA, HCA, mHC, flash at 8% cost of pro, which itself is 14% cost of opus), dropped the best open base models in the world, peaced out. BYO posttraining. leave that to the agent labs to pick up the scraps. bravo.
DeepSeek V4open modelsmodel efficiencyAI pricinglong context
Social Twitter Apr 29

👇 I was there for this, sitting right next to Altman. Realizing a few months later he lied about it ...

By @GaryMarcus

78 score
AI Analysis

Following yesterday's News coverage of the Musk v. OpenAI trial, Marcus reveals he was sitting next to Altman during sworn testimony where Altman lied by omission, says this turned him against Altman and led him to realize Altman is a 'serial liar'

Source details
👇 I was there for this, sitting right next to Altman. Realizing a few months later he lied about it (by omission) was what turned me against him. We had sworn to tell the whole truth and nothing but the truth. He didn’t. Soon after that, I realized that he was a serial liar. I have been trying to get the world to wake up to Sam’s lack of candor ever since.
musk_openai_lawsuitopenai_criticismaltman_credibility
Social Twitter Apr 29

terminal has been my primary interface to my computer for almost two decades. now it’s the Codex app...

By @gdb

78 score
AI Analysis

Adding to the Codex momentum from Social yesterday, Greg Brockman says the Codex app has replaced terminal as his primary computer interface after two decades

Source details
terminal has been my primary interface to my computer for almost two decades. now it’s the Codex app.
openai_codexdeveloper_toolsparadigm_shift
Social Twitter Apr 29

GPU library performance can be very notchy -- runtime of batched torch.linalg.solve_ex() went up by ...

By @ID_AA_Carmack

78 score
AI Analysis

John Carmack discovers GPU library performance is 'notchy' - batched torch.linalg.solve_ex() runtime jumps 10x going from 511x511 to 512x512 matrices

Source details
GPU library performance can be very notchy -- runtime of batched torch.linalg.solve_ex() went up by over 10x going from 511x511 matrices to 512x512.
gpu_performancepytorchcuda_optimizationtechnical_computingml_infrastructure
Social Twitter Apr 29

For years, voice AI has been stuck in a rigid loop: think, then speak. But real human conversation i...

By @hardmaru

75 score
AI Analysis

Hardmaru (David Ha) presents ICASSP 2026 research on tandem architecture for voice AI - 'speak while thinking' approach where fast speech model replies instantly while LLM injects knowledge in parallel

Source details
For years, voice AI has been stuck in a rigid loop: think, then speak. But real human conversation is messy, overlapping, and asynchronous. In our new #ICASSP2026 work, we built a tandem architecture that shifts the paradigm to “speak while thinking.” A fast speech model starts replying instantly, while a backend LLM runs in parallel to inject deep knowledge on the fly. It’s a completely different way to approach conversational AI, making it feel remarkably more alive. Blog: t.co/qZ4c
voice_airesearchconversational_aiarchitecture_innovation