Wiredframe Radar logo

Category intelligence

Social Media Briefing — May 28, 2026

467 current items analyzed and ranked.

Executive synthesis

Social Media Summary

Research breakthroughs led the conversation. hardmaru and Burkov spotlighted an ICLR 2026 paper reframing residual connections as diffusion denoising, enabling independently trained network blocks and major memory savings.

A persistent undercurrent across many posts was token economics—organizations blowing through budgets with unclear ROI, amplified by Ethan Mollick, Gary Marcus, and an MIT-led paper showing agents consume ~1000x more tokens than chat.

Key Themes

Training Efficiency Research · 3Token Economics and Enterprise AI Costs · 8AI Policy and Prosperity · 3AI Hiring Bias · 3Tokenization & Inference Performance · 6Continual Learning & RL · 1Neurosymbolic AI and AGI Roadmaps · 4Catastrophic Forgetting Research · 8Document Parsing & Open Source Tools · 3AI Agents and Autonomy · 9

Primary evidence

Top Ranked Signals

Social Twitter May 27

For over a decade, we’ve accepted that end-to-end backprop is the only way to train deep networks. B...

By @hardmaru

88 score
AI Analysis

Hardmaru announces an ICLR 2026 paper that breaks networks into independently trained blocks by treating the forward pass like diffusion denoising, slashing training memory while matching end-to-end performance on ViTs, DiTs, and LLMs.

For over a decade, we’ve accepted that end-to-end backprop is the only way to train deep networks. But holding the entire network in memory all at once is why AI training is hitting a resource wall. We found a new way to break the network into blocks and train them independently. The trick? Treating the network’s forward pass like a diffusion model denoising a signal. This reinterpretation slashes the memory needed to train deep models. In our #ICLR2026 paper (https://t.co/PK5h0mqQSo), we matc
training-efficiencydiffusion-modelsresearchbackpropagation-alternatives
Social Twitter May 27

AI should dramatically increase quality of life and individual freedoms for people around the world....

By @sama

80 score
AI Analysis

Sam Altman announces an initial $250M OpenAI Foundation commitment to measurement, transition support, and broadly shared prosperity efforts.

AI should dramatically increase quality of life and individual freedoms for people around the world. The OpenAI Foundation is making an initial $250M commitment to measurement, transition support, and new approaches to broadly shared prosperity. t.co/zOD8O94RjQ
openaiai-policyeconomic-impactphilanthropy
Social Twitter May 27

You should read this thread. It used to take about 25 seconds to generate a 5-second video on 8 Bla...

By @NVIDIAAI

74 score
AI Analysis

NVIDIA AI highlights HaoAI Lab cutting 5-second video generation from 25s on 8 Blackwell GPUs to 4.2s on a single Blackwell GPU, with the tech open sourced.

You should read this thread. It used to take about 25 seconds to generate a 5-second video on 8 Blackwell GPUs. The legends at @haoailab brought that down to just 4.2 seconds on a single Blackwell GPU… and then open sourced the tech behind it.
video generationGPU efficiencyopen sourceNVIDIA Blackwell
Social Twitter May 27

How much does a language model forget when finetuned on new tasks? We show both model size and optim...

By @andrewgwils

72 score
AI Analysis

Following yesterday's Research coverage, the authors break down their findings in a thread, Lead tweet of a research thread asking how much LMs forget when finetuned, showing model size and optimization matter and that self-generated replay nearly eliminates forgetting.

How much does a language model forget when finetuned on new tasks? We show both model size and optimization matter and forgetting can be nearly eliminated with self-generated replay! t.co/Qs9A4n095s w/@mrtnm @dongkyucho @ShikaiQiu @rumichunara @Pavel_Izmailov 1/8 t.co/Z4tTKGcnxA
catastrophic forgettingfinetuningML researchself-generated replay
Social Twitter May 27

We've created the world's fastest PDF parser ⚡️ And it's more accurate than any other open-source, ...

By @jerryjliu0

72 score
AI Analysis

jerryjliu announces LiteParse v2, a Rust-rewritten PDF parser claimed to be the fastest and most accurate model-free open-source parser, with Python and Node bindings supporting 50+ document types.

We've created the world's fastest PDF parser ⚡️ And it's more accurate than any other open-source, model-free PDF parser out there (pymupdf, pypdf, markitdown, pdftotext, opendataloader, pymupdf4llm) Introducing LiteParse v2 - we rewrote the entire library into Rust and adapted it as native packages for Python and Node. It supports 50+ different document types, can be triggered directly or installable directly within your favorite AI agent. Blog: t.co/ckb0G73ESs Repo: t.co/
document parsingopen sourceRustLlamaIndexproduct launch
Social Twitter May 27

Training a deep neural network with standard backpropagation requires keeping every layer's intermed...

By @burkov

70 score
AI Analysis

Burkov explains the ICLR 2026 paper that equates residual connections to diffusion denoising steps, allowing transformers to be partitioned into independently trained blocks to cut memory use.

Training a deep neural network with standard backpropagation requires keeping every layer's intermediate activations in memory until the backward pass uses them, so memory grows with depth and limits how large models can be made. Previous attempts to split a network into pieces that train independently have been ad-hoc, performed noticeably worse than ordinary training, or only worked on classification tasks. The authors of this ICLR 2026 paper notice that a residual connection (where a layer
training-efficiencydiffusion-modelsresearchtransformers
Social Twitter May 27

New research reveals alarming patterns in AI hiring tools. A large-scale study of 4 million job app...

By @StanfordHAI

68 score
AI Analysis

Following yesterday's Research coverage of algorithmic monocultures in hiring, Stanford HAI shares research on AI hiring tools showing algorithmic discrimination against 26% of Black and 15% of Asian applicants across 4M applications, amplified when one vendor serves many employers.

New research reveals alarming patterns in AI hiring tools. A large-scale study of 4 million job applications found that 26% of Black applicants and 15% of Asian applicants faced algorithmic discrimination. When one AI vendor screens for multiple employers, qualified candidates can be systematically rejected everywhere they apply. As 90% of U.S. employers now use AI screening, understanding these tools' impact on the workforce becomes more urgent. Hear more from the study's lead author @Rishi
AI biashiring algorithmsAI ethicsresearch
Social Twitter May 27

MIT just killed the idea that cheaper tokens mean cheaper AI. A paper from MIT, Stanford, Google De...

By @AlphaSignalAI

68 score
AI Analysis

AlphaSignal highlights a multi-institution paper finding AI agents consume roughly 1000x more tokens than chat, with cost driven by re-reading context each loop and 30x cost variance on identical tasks.

MIT just killed the idea that cheaper tokens mean cheaper AI. A paper from MIT, Stanford, Google DeepMind, and Microsoft found agents use 1000x more tokens than chat. The cost isn't the code they write. It's re-reading context on every loop. Same task, same model: 30x cost variance.
agentstoken economicsresearchcost optimization
Social Twitter May 27

We're open-sourcing the Unigram tokenizer we rebuilt to reduce CPU utilization by 5-6x. Small reran...

By @perplexity_ai

67 score
AI Analysis

Perplexity open-sources a rebuilt Unigram tokenizer that reduces CPU utilization 5-6x, addressing CPU tokenization as a major latency share for small rerankers and embedders.

We're open-sourcing the Unigram tokenizer we rebuilt to reduce CPU utilization by 5-6x. Small rerankers and embedders run in single-digit milliseconds on GPU, making CPU tokenization a meaningful share of total latency. t.co/QUnHeiho56 t.co/Oh29f1lo51
tokenizationopen sourcePerplexityperformance optimization
Social Twitter May 27

Introducing Dynamo Snapshot, our approach for fast startup for inference workloads on Kubernetes, wh...

By @NVIDIAAI

66 score
AI Analysis

NVIDIA introduces Dynamo Snapshot, reducing Kubernetes inference cold-start from minutes to under 5 seconds using GMS, Linux AIO, and parallel CRIU restore.

Introducing Dynamo Snapshot, our approach for fast startup for inference workloads on Kubernetes, which reduces startup time from minutes to under 5 seconds. In production inference deployments demand fluctuates over time. Cold-starting inference workloads can take minutes, leaving idle GPUs that generate no tokens and serve no requests. Snapshot leverages GMS to enable concurrent weight restoration over a high-speed interconnect, while using Linux native AIO and parallel memfd restoration to
NVIDIAinference infrastructureKubernetesGPU efficiency
Social Twitter May 27

If we had done everything I suggested in my 2020 arXiv article “The Next Decade in AI”, we might act...

By @GaryMarcus

65 score
AI Analysis

Gary Marcus revisits his 2020 essay The Next Decade in AI, arguing neurosymbolic progress has occurred via Claude Code and DeepMind math systems, but three other AGI goals remain unmet due to overcommitment to scaling.

If we had done everything I suggested in my 2020 arXiv article “The Next Decade in AI”, we might actually have reached AGI by now. In the last three years, after a detour driven by the false promise of pure scaling, we have gone a long way to the first of the four goals I laid out: neurosymbolic AI. (Claude Code; DeepMind’s new math system that solved 9 Erdos problems, etc). Harnesses and tools, for example. But the field has made little progress on the other three:
  • databases of explicit, sy
agineurosymbolic-aiscaling-debatereasoning
Social Twitter May 27

Introducing Runway MCP. Now you can connect Runway directly into Claude, ChatGPT, Cursor, Replit and...

By @runwayml

65 score
AI Analysis

Runway introduces Runway MCP, letting users connect Runway image/video models (Gen-4.5, Seedance 2.0, GPT Images 2.0, Kling) directly into Claude, ChatGPT, Cursor, and Replit.

Introducing Runway MCP. Now you can connect Runway directly into Claude, ChatGPT, Cursor, Replit and more. Generate polished images and videos with state-of-the-art models, like Gen-4.5, Seedance 2.0, GPT Images 2.0, Kling and more. Right from where you're already working. Connect the MCP to your agent in seconds at the link below.
RunwayMCPmedia generationproduct launch