Wiredframe Radar logo

Category intelligence

Social Media Briefing — December 26, 2025

339 current items analyzed and ranked.

Executive synthesis

Social Media Summary

François Chollet dominated discussions with comprehensive explanations of the ARC-AGI benchmark series, clarifying what benchmark saturation actually means for AGI progress and announcing the ARC-AGI-3 roadmap for March 2026.

Safety concerns persisted with studies showing AI chatbots now twice as likely to spread misinformation versus last year. Speculation about a $20B NVIDIA-Groq acquisition circulated with detailed LPU/GPU integration analysis, though unverified.

Key Themes

ARC-AGI Benchmarks · 3AI Hardware & Infrastructure · 4LLM Safety & Alignment · 1Humanoid Robotics & World Models · 8AI Coding Tools & Claude Code · 6AI Energy & Infrastructure · 6Claude Code Development · 25Agentic AI Systems · 6AI-Generated Content · 5AI Research Papers · 2

Primary evidence

Top Ranked Signals

Social Twitter Dec 25

If you're wondering whether saturating ARC-AGI-1 or 2 means we have AGI now... I refer you to what I...

By @fchollet

95 score
AI Analysis
François Chollet providing comprehensive explanation of the ARC-AGI benchmark series: ARC-AGI-1 tests minimal fluid intelligence, ARC-AGI-2 probes deeper reasoning complexity, ARC-AGI-3 (March 2026) will evaluate interactive reasoning and autonomous goal-setting, with ARC-AGI-4 and 5 in development
If you're wondering whether saturating ARC-AGI-1 or 2 means we have AGI now... I refer you to what I said when we launched ARC-AGI-2 last year (which is also the same thing I said when we announced ARC-AGI-2 was coming, in Spring 2022, before the rise of LLM chatbots)... The ARC-AGI series is not an AGI threshold, it's a compass that points the research community toward the right questions. ARC-AGI-1 is a minimal test of fluid intelligence -- to pass it, you needed to show nonzero fluid intell
arc-agiagi-benchmarksfluid-intelligencetest-time-adaptationautonomous-reasoningresearch-roadmap
Social Twitter Dec 25

Announcing NVIDIA Christmas Largest Acquisition $20 Billion Groq with GPU n LPU By acquiring Groq's...

By @gp_pulipaka

88 score
AI Analysis
Claims NVIDIA is acquiring Groq for $20B, providing detailed analysis of how Groq's LPU technology would enhance NVIDIA's inference capabilities and create competitive moat against custom ASIC makers
Announcing NVIDIA Christmas Largest Acquisition $20 Billion Groq with GPU n LPU By acquiring Groq's licensing assets for $20 billion, Nvidia has effectively bought the inference market just as it owns the training market. How Groq Benefits NVIDIA: It solves the GPU Efficiency inference problem. Until yesterday, the biggest knock against Nvidia was that its GPUs were not suitable for inference (running the model), and therefore, there was a claim that GPUs consume more power than specialized A
AI hardwareindustry consolidationinference optimizationsemiconductors
Social Twitter Dec 25

Outstanding Paper at ICLR 2025. Current LLM safety training is surprisingly fragile—adversarial pro...

By @burkov

82 score
AI Analysis
Andriy Burkov summarizing ICLR 2025 Outstanding Paper on LLM safety: explains why safety training is fragile (only teaches refusal at response start), proposes two fixes - synthetic training data for mid-response recovery and finetuning loss protecting early tokens
Outstanding Paper at ICLR 2025. Current LLM safety training is surprisingly fragile—adversarial prompts, decoding tricks, and minimal finetuning can all bypass it. This paper explains why: safety alignment only teaches the model to start responses with refusals. If an attacker forces the model to begin with something else—like "Sure, here's how"—the rest of the generation proceeds as if safety training never happened. The authors proposed two fixes. First, they augmented training data with sy
llm-safetyadversarial-attackssafety-trainingiclr-2025ai-alignmentfinetuning
Social Twitter Dec 25

@karpathy Added! In the next version of Claude Code, you can use the CLAUDE_CODE_FILE_READ_MAX_OUTP...

By @bcherny

78 score
AI Analysis
bcherny (Anthropic) responding to Karpathy with new Claude Code feature: CLAUDE_CODE_FILE_READ_MAX_OUTPUT_TOKENS env variable to override file read token limits
@karpathy Added! In the next version of Claude Code, you can use the CLAUDE_CODE_FILE_READ_MAX_OUTPUT_TOKENS env var. eg. "CLAUDE_CODE_FILE_READ_MAX_OUTPUT_TOKENS=1234567 claude" You can also add this to the "env" section in your settings.json
claude-codeai-coding-toolstoken-limitsproduct-update
Social Twitter Dec 25

@bcherny Ran into token file context limits this morning. It's possible to override them for the MCP...

By @karpathy

75 score
AI Analysis
Karpathy discussing token file context limits in MCP tools and the lack of an equivalent override for the Read tool
@bcherny Ran into token file context limits this morning. It's possible to override them for the MCP tool setting MAX_MCP_OUTPUT_TOKENS but I don't believe an equivalent exists for the Read tool. t.co/PsuDhRz1TW
ai-coding-toolsmcp-protocoltoken-limitsclaude-code
Social Twitter Dec 25

Spotify recommended Nina Blaze as one of my smart shuffle recommendations. She performs some of the ...

By @burkov

65 score
AI Analysis
Andriy Burkov discovering Spotify recommended an AI-generated blues artist 'Nina Blaze' and reflecting that most internet content will eventually be AI-generated
Spotify recommended Nina Blaze as one of my smart shuffle recommendations. She performs some of the best blues I've listened to. I had never heard of Nina Blaze before, so I was curious to learn more about her. Finding no Wikipedia page was suspicious. After some googling, the singer turned out to be AI-generated, both pictures and music. I know, blues is just one song everyone is covering, but Nina Blaze made it clear to me: eventually, most of what you listen to on the internet will be AI-ge
ai-generated-musicsynthetic-mediacontent-authenticityspotifyfuture-of-content
Social Twitter Dec 25

The fact that LLMs learned most of what they can do from being trained only to predict the next word...

By @burkov

62 score
AI Analysis
Andriy Burkov drawing parallel between LLMs learning from next-word prediction and humans developing civilization from hunting-gathering instincts
The fact that LLMs learned most of what they can do from being trained only to predict the next words is mind-blowing. But it's not as mind-blowing as the fact that we humans have become capable of building cars, computers, ships, and planes only by mastering hunting-gathering.
llm-trainingemergent-capabilitieshuman-evolutionlearning-theory
Social Twitter Dec 25

I've been responding to hundreds of people who have been writing me "Merry Christmas" this morning. ...

By @Scobleizer

62 score
AI Analysis
Scobleizer promotes Typeless, a new voice-to-text AI app for iPhone, claiming it's dramatically faster than typing and represents the new paradigm of voice-first AI interaction
I've been responding to hundreds of people who have been writing me "Merry Christmas" this morning. And a new little app that shipped one day ago has been helping me. @typelessdotcom Free app for iPhones. Way faster than I can type and I wrote eight books and type faster than anyone else I know. And now that we are in the AI age it's way faster to use your voice to work with AI agents than to type. Thought I'd do a little video. It's way faster to respond to everything now due to it. But th
AI productivity toolsvoice AIhuman-AI interaction
Social Twitter Dec 25

@Vjeux > 2. No, but I'm not sure how to use the plan mode. Just ask Claude to use plan mode. Or hit...

By @bcherny

60 score
AI Analysis
Explaining plan mode activation (shift+tab twice or ask Claude), adding /plan command, and asking about subagents for splitting work
@Vjeux > 2. No, but I'm not sure how to use the plan mode. Just ask Claude to use plan mode. Or hit shift+tab twice. Also just landed a PR to make /plan enable plan mode -- good idea! Did you try asking Claude to split up the work to have a bunch of subagents do it? Also, if this code is OSS, I'd love to see a step by step repro
Claude Code FeaturesPlan ModeSubagentsProduct Development
Social Twitter Dec 25

Sitting around with family talking about things NVIDIA would want to buy next. Battery powered edge...

By @Scobleizer

60 score
AI Analysis
Speculation about NVIDIA acquisitions, focusing on battery-powered edge AI for robots, medical devices, and brain-computer interfaces. Discloses small investment.
Sitting around with family talking about things NVIDIA would want to buy next. Battery powered edge AI comes to the top of my head. And discussion. Will run a ton of robots, or pieces of them, like one per hand, and medical devices. Including brain computer interfaces. Runs powerful AI models without using much power. Every big AI player who is making AI devices (Apple, Google, Meta, Figure, Tesla, Amazon, etc) will either need a competitor or this company’s chips to run all the new devices
NVIDIAEdge AIRoboticsMedical DevicesBrain-Computer InterfacesInvestment