Category intelligence

Social Media Briefing — April 11, 2026

527 current items analyzed and ranked.

Executive synthesis

Social Media Summary

Anthropic's `/ultraplan` for Claude Code dominated the day with 703K views, introducing cloud-based planning separated from local implementation—a meaningful architectural shift in coding agents. Sam Altman published a rare personal blog post generating 2.3M views, while user frustration with Claude Code's overeager behavior also trended.

Ethan Mollick delivered a comprehensive assessment of frontier AI: Google, OpenAI, and Anthropic lead decisively, possibly exhibiting recursive self-improvement; Chinese models trail 7–9 months behind; open-weights frontier development is fading
Andrej Karpathy sparked viral discussion framing detailed LLM finetuning on personal interviews as a tractable 'brain uploading'
NVIDIA's EGGROLL research showed evolution strategies can rival backpropagation at 14B parameters, using integer-only math with 100x training throughput gains
Allen AI open-sourced the full MolmoWeb stack for web agents; Hugging Face CEO Clément Delangue announced Kernels, a new repo type for hardware-optimized operations
Harrison Chase argued agent harnesses represent the first stable agent abstraction, while François Chollet offered a deep take on symmetry as compression in physics and intelligence

Key Themes

Claude Code Features (Ultraplan) · 4State of Frontier AI Competition · 7Alternative Training Paradigms (EGGROLL/Evolution Strategies) · 2LLM-Based Digital Identity & Brain Upload · 3Claude/Anthropic Product Quality & Direction · 5Agent Architecture & Harnesses · 8Open Source Web Agents (MolmoWeb) · 5LLM Capabilities & Limitations · 8AI Agent Architecture & Tools · 12Voice AI Progress · 2

Primary evidence

Top Ranked Signals

Social Twitter Apr 10

New in Claude Code: /ultraplan Claude builds an implementation plan for you on the web. You can rea...

By @trq212

95 score

AI Analysis

Anthropic's trq212 announces /ultraplan for Claude Code: a new feature where Claude builds an implementation plan on the web that users can read, edit, and execute either on the web or in terminal. Available in preview for all Claude Code web users.

New in Claude Code: /ultraplan Claude builds an implementation plan for you on the web. You can read it and edit it, then run the plan on the web or back in your terminal. Available now in preview for all users with CC on the web enabled.

claude_code_featuresagentic_codingproduct_launches

Social Twitter Apr 10

I wrote this early this morning and I wasn't sure if I would actually publish it, but here it is: h...

By @sama

88 score

AI Analysis

Sam Altman published a personal blog post he was hesitant to share, generating massive engagement (2.3M views). Content not visible but the hesitation and scale suggest something significant/vulnerable.

I wrote this early this morning and I wasn't sure if I would actually publish it, but here it is: t.co/7Dw9UFpeep

OpenAI leadershipindustry narrative

Social Twitter Apr 10

@jenzhuscott Yes it's the tractable form of brain upload. There's a ton of scifi on brain uploads th...

By @karpathy

82 score

AI Analysis

Karpathy describes 'brain upload via LLM' as the tractable form of brain uploading - detailed video interviews plus LLM finetuning to create a simulation of a person with their knowledge and personality. Points to HeyGen as early example. 226K views.

@jenzhuscott Yes it's the tractable form of brain upload. There's a ton of scifi on brain uploads that requires way too exotic tech (scanning and simulating brains etc), when we're about to get a lossy and approximate version of that *a lot* sooner via LLM simulators. You can easily imagine a "brain upload" startup - you show up for a few days to carry out detailed video interviews, then they use all that data with an LLM finetuning process to "upload" you and give you an API endpoint of your si

LLM simulationdigital identityfuture of AIproduct concepts

Social Twitter Apr 10

Our latest Live model is # 1 on Tau Voice Bench! Excited to see this new frontier of voice models ...

By @OfficialLoganK

78 score

AI Analysis

Google's Logan (OfficialLoganK) announces their latest Gemini Live model is #1 on Tau Voice Bench, marking progress in voice model usability in production.

Our latest Live model is # 1 on Tau Voice Bench! Excited to see this new frontier of voice models cross the chasm of usability in production. t.co/wKphNSV6SL

voice_aigoogle_aibenchmarks

Social Bluesky Apr 10

So we now have a pretty good picture of the state of the frontier AI model makers. 1/ US closed sou...

By @emollick.bsky.social

78 score

AI Analysis

As first posted on Social yesterday, Emollick's thread opener assessing the state of frontier AI: Google, OpenAI, and Anthropic lead well ahead of all others, and may have some form of recursive self-improvement operating.

So we now have a pretty good picture of the state of the frontier AI model makers. 1/ US closed source models continue to lead. Google, OpenAI, and Anthropic stand well ahead of every other lab, and may have some form of recursive self-improvement operating.

frontier AIrecursive self-improvementGoogleOpenAIAnthropicAI competitionindustry analysis

Social Twitter Apr 10

NVIDIA just trained a 14-billion-parameter AI using evolution, not calculus. Every AI today learns...

By @AlphaSignalAI

75 score

AI Analysis

AlphaSignalAI describes NVIDIA's EGGROLL: training a 14B parameter AI model using evolution strategies instead of backpropagation. Uses low-rank matrix decomposition for mutations, achieving 100x faster training throughput, 91% inference speed, competitive with backprop on reasoning, works with integer-only computation.

NVIDIA just trained a 14-billion-parameter AI using evolution, not calculus. Every AI today learns through backpropagation. It computes gradients, adjusts weights, repeats. It works, but it demands precision hardware and enormous GPU clusters. Evolution Strategies offered an alternative. Mutate the model, test it, keep what works. Like biological evolution. The problem was speed. Random mutations on GPUs were painfully slow. EGGROLL fixes this with one trick. It splits huge rando

evolution_strategiesalternative_trainingnvidia_researchml_researchbackpropagation_alternatives

Social Twitter Apr 10

You can now train, adapt, and eval web agents on your own tasks. We're releasing the full MolmoWeb...

By @allen_ai

75 score

AI Analysis

Allen AI releases the full MolmoWeb codebase—training code, eval harness, annotation tooling, synthetic data pipeline, and client-side demo code—for building and adapting web agents.

You can now train, adapt, and eval web agents on your own tasks. We're releasing the full MolmoWeb codebase—the training code, eval harness, annotation tooling, synthetic data pipeline, & client-side code for our demo. 🧵 t.co/yMGRuzbeXQ

open_source_aiweb_agentsresearch_releases

Social Twitter Apr 10

im excited about agent harnesses because i think are the first stable agent abstractions we can buil...

By @hwchase17

73 score

AI Analysis

Following yesterday's News on LangChain's Deep Agents Deploy launch, Harrison Chase (LangChain CEO) argues agent harnesses are the first stable agent abstractions. Models are now good enough that 'running LLMs in a loop calling tools' actually works, unlike the early AutoGPT days. He sees deepagents as a stable foundation rather than the shifting frameworks of the past.

im excited about agent harnesses because i think are the first stable agent abstractions we can build on top (which is why we're investing so much in deepagents) we always wanted to run llms in a loop and have them call tools (remember autoGPT? that's all that was) but the models weren't good enough so we built chains and other architectures as a proxy as the models got better, the "right" way to build the most agentic systems changed pretty dramatically which meant the frameworks (like langc

agent_architectureai_frameworksindustry_reflection

Social Twitter Apr 10

In a world where writing code to build websites and apps is trivial (thank you Lovable, Cursor, Clau...

By @ClementDelangue

72 score

AI Analysis

Clement Delangue announces Hugging Face is launching 'Kernels' - a new repo type for optimized binary operations supporting CUDA, ROCm, Apple Silicon, and Intel XPU. Positions it as enabling AI builders (not just users) in a world where code generation is commoditized.

In a world where writing code to build websites and apps is trivial (thank you Lovable, Cursor, Claude,...), the real differentiation for you and your company (and what makes you successful) will be how you manage to train, run and optimize AI models yourself. That's why at Hugging Face, we're doubling down on enabling more to become AI builders rather than AI users. We're releasing this week Kernels on the Hugging Face hub. This repo type is for the hardcore AI engineers among you. Kernels ar

Hugging FaceAI infrastructurehardware optimizationopen source AIdeveloper tools

Social Twitter Apr 10

I canceled Claude Code but still have to use it for 2 weeks because it was paid. OMG, it's so dumb! ...

By @burkov

72 score

AI Analysis

Burkov vents frustration with Claude Code's behavior of immediately acting instead of answering questions, announces he cancelled it, and hopes Claude-Mythos will restore the quality of Claude Opus 4.4.

I canceled Claude Code but still have to use it for 2 weeks because it was paid. OMG, it's so dumb! You say, "Why did you do X?" And it's like: "You are right! I shouldn't have done X. I'm now doing Y!" And you're like, "ANSWER THE FUCKING QUESTION!" and it's like, "You are right to push back. I didn't think it through. Now I'm implementing solution Z!" OMFG! If this Mythos will bring the original Claude Opus 4.4 back, it will already be a success for Anthropic.

claude_code_criticismanthropic_product_qualityclaude_mythosai_coding_tools

Social Twitter Apr 10

A few questions on plan vs ultraplan. I think the insight here is that implementation can sometime...

By @trq212

72 score

AI Analysis

trq212 shares insight that planning can happen in the cloud since it's about reading code and understanding intent, while implementation may require local environment/interactivity—the key architectural distinction behind plan vs ultraplan.

A few questions on plan vs ultraplan. I think the insight here is that implementation can sometimes require a local environment/interactivity, but planning can happen in the cloud since it's mostly about reading code & understanding intent.

claude_code_featuresagentic_codingagent_architecture

Social Bluesky Apr 10

Lots of open questions: is RSI indeed happening at the Big Three labs? How long can the exponential ...

By @emollick.bsky.social

72 score

AI Analysis

As first posted on Social yesterday, Emollick raises key open questions: Is RSI happening at Big Three labs? How long can LLM exponential gains last? Can Chinese models keep pace with chip constraints? Will frontier open-weights models continue?

Lots of open questions: is RSI indeed happening at the Big Three labs? How long can the exponential of LLM ability gain last? How much do Chinese models rely on distillation, and can they keep pace given chip constraints? Will there continue to be frontier open weights models?

recursive self-improvementAI scalingChina AIopen weightsfrontier models