Wiredframe Radar logo

Category intelligence

Social Media Briefing — January 23, 2026

512 current items analyzed and ranked.

Executive synthesis

Social Media Summary

AI business and infrastructure dominated discussions. Sam Altman revealed OpenAI added over $1B ARR in a single month from APIs alone, highlighting the often-overlooked enterprise side. The vLLM creators announced Inferact with a massive $150M seed led by a16z, signaling major investment in open-source inference infrastructure.

DeepMind announced D4RT for 4D video reconstruction (18x-300x faster). Ethan Mollick highlighted research showing humans increasingly using "ChatGPT-favorite words" in spoken language—"model collapse, except for humans." Anthropic shipped the Tasks upgrade for longer Claude projects, while OpenAI's Logan publicly praised Gemini 3 Flash as "highly underrated."

Key Themes

AI Infrastructure & Inference · 12AI Business & Economics · 8AGI Timeline Signals · 3Model Capabilities & Benchmarks · 7Claude Product Updates · 6Video Generation Milestones · 5Reinforcement Learning Research · 3Humanoid & Home Robots Economics · 8Research Advances · 8Open Source AI & Training Data Ethics · 6

Primary evidence

Top Ranked Signals

Social Twitter Jan 22

We have added more than $1B of ARR in the last month just from our API business. People think of us...

By @sama

95 score
AI Analysis

Sam Altman announces OpenAI added more than $1B in ARR in the last month from API business alone, highlighting the often-overlooked enterprise side of OpenAI beyond ChatGPT.

We have added more than $1B of ARR in the last month just from our API business. People think of us mostly as ChatGPT, but the API team is doing amazing work!
AI BusinessOpenAIEnterprise AI
Social Twitter Jan 22

Today, we're proud to announce @inferact, a startup founded by creators and core maintainers of @vll...

By @woosuk_k

95 score
AI Analysis

Major announcement: Inferact, startup founded by vLLM creators (woosuk_k, simon_mo_, KaichaoYou, others), announces $150M seed round led by a16z and Lightspeed. Mission is to grow vLLM as world's AI inference engine. vLLM supports 500+ model architectures, 200+ accelerator types, with 2000+ contributors.

Today, we're proud to announce @inferact, a startup founded by creators and core maintainers of @vllm_project, the most popular open-source LLM inference engine. Our mission is to grow vLLM as the world's AI inference engine and accelerate AI progress by making inference cheaper and faster. The Challenge Inference is not solved. It's getting harder. Models grow larger. New architectures proliferate: mixture-of-experts, multimodal, agentic. Every breakthrough demands new infrastructure. Meanwh
AI infrastructurestartup fundingopen source AILLM inference
Social Twitter Jan 22

AGI is now on the horizon and it will deeply transform many things, including the economy. I'm curr...

By @ShaneLegg

94 score
AI Analysis

Shane Legg (DeepMind co-founder) declares 'AGI is now on the horizon' and is hiring a Senior Economist to lead a team investigating post-AGI economics.

AGI is now on the horizon and it will deeply transform many things, including the economy. I'm currently looking to hire a Senior Economist, reporting directly to me, to lead a small team investigating post-AGI economics. Job spec and application here: t.co/VAfwrMc8Tp
AGI TimelineDeepMindAI EconomicsAI Safety
Social Twitter Jan 22

New on the Anthropic Engineering Blog: We give prospective performance engineering candidates a noto...

By @AnthropicAI

92 score
AI Analysis

Anthropic reveals their notoriously difficult take-home exam for performance engineering candidates was beaten by Claude Opus 4.5, forcing a redesign.

New on the Anthropic Engineering Blog: We give prospective performance engineering candidates a notoriously difficult take-home exam. It worked well—until Opus 4.5 beat it. Here's how we designed (and redesigned) it: t.co/3RZVyhpVij
Model CapabilitiesAnthropicBenchmarksClaude
Social Twitter Jan 22

#PaperADay 9 floq: Training Critics via Flow-Matching for Scaling Compute in Value-Based RL https://...

By @ID_AA_Carmack

92 score
AI Analysis

John Carmack's #PaperADay analysis of 'floq' paper on using flow-matching for value-based RL. Provides deep technical commentary on using flow models to generate categorical values in ~8 steps, discusses potential for error correction, compares against ensembles and deeper networks, and notes key implementation details like Fourier basis for time input.

#PaperADay 9 floq: Training Critics via Flow-Matching for Scaling Compute in Value-Based RL t.co/V4kCMfUkEK In theory, value based reinforcement learning is a regression problem, which is most naturally addressed with an MSE loss. However, there are a bunch of subtle reasons why a neural network terminating in a single activation may not be an ideal value function approximator. Most models today use a categorical value representation to mitigate some of the training issues, but the pro
reinforcement learningflow modelstechnical researchvalue functions
Social Twitter Jan 22

Everyone is starting to sound like AI, even in spoken language Analysis of 280,000 transcripts of v...

By @emollick

88 score
AI Analysis

Analysis of 280,000 video transcripts shows humans increasingly using ChatGPT-favorite words in spoken language. Emollick calls it 'model collapse, except for humans.'

Everyone is starting to sound like AI, even in spoken language Analysis of 280,000 transcripts of videos of talks & presentations from academic channels finds they increasingly used words that are favorites of ChatGPT Model collapse, except for humans t.co/nlfbEozhCJ t.co/POW0cf17G6
AI-Human InteractionCultural ImpactLanguageResearch
Social Twitter Jan 22

We're helping AI to see the 3D world in motion as humans do. 🌐 Enter D4RT: a unified model that tur...

By @GoogleDeepMind

87 score
AI Analysis

Google DeepMind announces D4RT, a unified model for 4D video reconstruction that runs 18x-300x faster than previous methods, processing 1-minute videos in ~5 seconds.

We're helping AI to see the 3D world in motion as humans do. 🌐 Enter D4RT: a unified model that turns video into 4D representations faster than previous methods - enabling it to understand space and time. This is how it works 🧵
ResearchDeepMindComputer VisionVideo Understanding
Social Twitter Jan 22

GPT-5.2 Pro is good enough to check reproducibility & robustness of academic papers across many ...

By @emollick

85 score
AI Analysis

Emollick claims GPT-5.2 Pro is capable enough to check reproducibility and robustness of academic papers across many fields, though not yet able to do independent replication.

GPT-5.2 Pro is good enough to check reproducibility & robustness of academic papers across many fields (given the data, can you get the same results? are the statistics brittle?). At scale, this would have a big impact. It can't do an independent replication with new data, yet.
Model CapabilitiesGPT-5.2Scientific ResearchReproducibility
Social Twitter Jan 22

We have been thinking a lot about what happens when generated and non-generated content are indistin...

By @c_valenzuelab

85 score
AI Analysis

Runway CEO reports study finding over 90% of participants couldn't reliably distinguish Gen-4.5 outputs from real video - 'tipping point' crossed

We have been thinking a lot about what happens when generated and non-generated content are indistinguishable from one another. This is the first study to try to see if that threshold has been crossed. We have officially reached a tipping point: over 90% of participants could not reliably distinguish Gen-4.5 outputs from real video. We still have a lot to figure out collectively and culturally to make the best out of what’s coming ahead.
video_generationhuman_indistinguishabilitygen4ai_milestonescontent_authenticity
Social Twitter Jan 22

Given the attention to Claude Code/Codex, I think that people's views about what AI can or can't do ...

By @emollick

82 score
AI Analysis

Emollick argues that current perceptions of AI capabilities are overly shaped by CLI tools like Claude Code/Codex, and different agentic harnesses will reveal different capability profiles.

Given the attention to Claude Code/Codex, I think that people's views about what AI can or can't do are getting overly shaped by the affordances of CLI tools. A different agentic harness will radically change the ability profile of frontier models We just haven't seen them yet.
Agentic AIAI ToolsModel CapabilitiesUX Design
Social Twitter Jan 22

Yesterday, I spoke with a developer who stopped contributing to open-source projects. 10+ years doi...

By @svpino

82 score
AI Analysis

Santiago shares story of developer who stopped open-source contributions after 10+ years because they don't want their work to become AI training data. High engagement (2200+ likes) reflects widespread concern.

Yesterday, I spoke with a developer who stopped contributing to open-source projects. 10+ years doing this, and now he's done. "Why should I spend my weekends building something that will just become training data for the next model?" I don't like where this is going.
AI ethicsopen sourceAI training datadeveloper sentiment