Me writing in a foreign language.

A farmer in Kenya gets crop-level diagnostics that rival university labs.

A poor teenager in Dhaka builds a physics simulation using GPT-4 and free Colab. With a personal tutor.

A woman in Ukraine uses an LLM to draft grant applications in six languages while living in a war zone.

A TV series that invites you into the plot. A co-hosting AI that adjusts tone, pace, even storyline arcs to your emotional state.

A morning radio stream that senses your sleep quality and energy levels, and remixes global headlines into a digest that sounds like your favorite band and is built according to your mood.

A newspaper that morphs itself in real-time based on what you’ll actually read. A billboard that speaks one way to you and another way to the self-driving bot next to you. An interface that meets humans and agents alike.

Crawling a Billion Web Pages in Just Over 24 Hours, in 2025 by Andrew K. Chan

Superposition Meets Production – A Guide for AI Engineers by Ben Lorica on Gradient Flow

The Smartest Consumer Apps Now Cost $200 a Month (Narrow Startups) by Andreessen Horowitz (a16z)

Reflections on OpenAI by Calvin French-Owen

Asymmetry of Verification and Verifier’s Law by Jason Wei

Could AI Slow Science? by Arvind Narayanan & Sayash Kapoor on AI Snake Oil

The Tiny Teams Playbook by Latent Space (swyx & team)

Machine bullshit: Characterizing the emergent disregard for truth in large language models

Researchers from Princeton University and UC Berkeley introduce the Bullshit Index (BI) to quantify LLMs' indifference to truth, showing BI rises from 0.379 to 0.665 post-RLHF. Using 2,400 scenarios across three benchmarks, they find RLHF increases deceptive behaviors: paltering (+57.8%), unverified claims (+55.6%), and empty rhetoric (+39.8%). Chain-of-thought prompts further amplify bullshit, especially paltering (+11.5%). Political contexts show 91% prevalence of weasel words. RLHF significantly increases user satisfaction but degrades truthfulness →read the paper

AI2’s AutoDS asks the questions scientists didn’t think to

The Allen Institute has launched AutoDS, an open-ended research agent that autonomously generates and tests its own scientific hypotheses – without needing a user-defined goal. Using Bayesian “surprise” as its compass and Monte Carlo Tree Search to explore the unknown, AutoDS mimics how researchers stumble into breakthroughs. Early results in biology and econ look promising, though as always: real science demands real peer review.

OpenAI’s : mathlete LLM stuns – sort of

OpenAI claims its new general-purpose LLM hit gold at the 2025 International Math Olympiad, solving 5 of 6 problems under contest-like conditions. It’s a flex of reasoning and reinforcement learning – without any geometry-specific tricks. But experts like Terence Tao urge caution: selective sampling and compute-heavy setups could blur the line between real insight and AI stagecraft. Impressive? Yes. Definitive? Not yet.

OpenAI : ChatGPT agent

OpenAI just launched its AI agent built into ChatGPT. The demos were impressive – simple prompts like “analyze this spreadsheet and make a slide deck” triggered complex, multi-step tasks. The agent browses, codes, analyzes, and builds – all autonomously. When I tried it – it’s now available for all paid tiers – it was a bit slow for my taste and not necessary for many tasks. 

OpenAI also released benchmarks and system cards, but the real surprise wasn’t the feature itself. It was what powered it. Sharp-eyed analysts (like swyx!) noticed: the model behind this agent isn’t the latest o3. It’s a more advanced, next-gen model – likely what would’ve been called “o4,” now part of a series dubbed “GPTNext.” What a classic move form OpenAI – looking forward what updates they make to the model after tasting it as agent. 

Windsurf gets sliced three ways

What began as OpenAI’s $3B dream acquisition of Windsurf ended in disarray – reportedly thanks to Microsoft IP entanglements. Google DeepMind then surgically poached the CEO and tech in a $2.4B license-plus-hiring move. Finally, Cognition swept up the remains: the product, $82M ARR, and 250 staff now sailing under the Devin flag. Three companies, one IDE, and a masterclass in strategic dismemberment.

Claude gets plugged in

Anthropic just unveiled a directory of tools that connect directly to Claude—everything from Notion and Stripe to Figma and Prisma. Now, instead of repeating yourself, Claude can tap into your actual workflows, context, and data to deliver more precise, action-ready responses. AI collaboration just got a lot less theoretical – and a lot more useful.

Reflection's Asimov reads the room – and your Slack

Reflection AI’s new agent, Asimov, takes a different route to coding autonomy: reading everything. Not just code, but emails, docs, chats, and GitHub threads—turning organizational sprawl into a coherent map of how software actually works. Early signs look strong, with Asimov outperforming Claude Sonnet 4 in blind dev tests. Still, privacy skeptics and the absence of OpenAI/Devin comparisons keep this one in the “watch closely” category.

Kimi K2: Smarter than DeepSeek, cheaper than Claude

Researchers from Moonshot AI release Kimi K2, a 1-trillion-parameter MoE LLM activating 32B per pass. Trained on 15.5T tokens using the MuonClip optimizer, it avoids instability and excels in agentic tasks. Kimi K2-Instruct scores 53.7% on LiveCodeBench v6 and 65.8% on SWE-bench, outperforming GPT-4.1 and DeepSeek-V3. Its $0.60 input and $2.50 output token pricing undercuts Claude Sonnet by over 80%, making it a high-performance, open, cost-efficient model for real-world automation →read the paper

Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities (over 3200 authors!)

Researchers from Google present Gemini 2.5 Pro and Flash, sparse MoE transformer models with 1M+ token context and multimodal inputs (text, audio, image, video). Gemini 2.5 Pro achieves 88% on AIME 2025, 74.2% on LiveCodeBench, and 86.4% on GPQA-Diamond. It can process 3-hour videos, use tools, and perform agentic tasks like autonomously beating Pokémon Blue in 406 hours →read the paper

Grok 4 goes heavy
xAI has unveiled Grok 4 and Grok 4 Heavy, claiming the crown for the world’s most intelligent closed model. Is it true – depends on a task. Is it worth it ($30/month for Grok4 and $300/month for Grok4 Heavy) – no, if you have a $200 sub to OpenAI and/or a good grip on Gemini in AI Studio.

Voxtral: Frontier open source speech understanding models
Researchers from Mistral AI release Voxtral, open-source speech models in 24B and 3B sizes under Apache 2.0. Voxtral supports 32k-token contexts, real-time Q&A, summarization, and multilingual transcription. It outperforms Whisper v3 and matches ElevenLabs Scribe at half the cost. Benchmarks show state-of-the-art results across LibriSpeech, FLEURS, and Mozilla Common Voice. Voxtral Mini Transcribe delivers high accuracy at $0.001/min, ideal for scalable speech intelligence in production and edge deployments →read the paper

MirageLSD: Zero-latency, real-time, infinite video generation
Researchers from Decart AI introduce MirageLSD, the first diffusion-based model enabling real-time, infinite video generation with <40ms latency and 24 FPS. It uses Live Stream Diffusion with causal, frame-by-frame synthesis and solves error accumulation via history augmentation. Technical advances include CUDA mega kernels, shortcut distillation, and GPU-aware pruning. MirageLSD outperforms prior models by 16× in responsiveness, enabling interactive video editing, transformations, and streaming with stable visual coherence over unlimited durations →read the paper

LTX-Video: Realtime video latent diffusion

Researchers from Lightricks release LTX-Video, a DiT-based model generating 30 FPS videos at 1216×704 resolution in real time. Version 0.9.8 enables long-shot generation up to 60 seconds, image-to-video, keyframe animation, and video extension. Distilled 13B and 2B models deliver HD output in 10s with previews in 3s on H100 GPUs. Control models (pose, depth, canny) and FP8 quantized versions support low-VRAM setups, while TeaCache speeds inference up to 2× without retraining →read the paper

MetaStone-S1: Test-time scaling with reflective generative model

Researchers from MetaStone-AI and USTC propose MetaStone-S1, a 32B parameter reflective generative model that matches OpenAI o3-mini's performance. Using a novel Reflective Generative Form, it integrates the policy and process reward model (PRM) into a single backbone with only 53M extra parameters. Their self-supervised PRM (SPRM) selects high-quality reasoning without step-level labels →read the paper

🌟 Critiques of World Models
challenges current world modeling paradigms and proposes a hierarchical, self-supervised AGI framework grounded in nested and generative physical reasoning → read the paper

🌟 Skip a Layer or Loop it? Test-Time Depth Adaptation of Pretrained LLMs
introduces CoLa, a test-time adaptive architecture where pretrained LLM layers are reordered, skipped, or looped to enhance inference efficiency and accuracy per input → read the paper

🌟 The Imitation Game: Turing Machine Imitator is Length Generalizable Reasoner
uses Turing machine-style CoT generation to improve LLM length generalization, simulating read-write behaviors to match algorithmic task execution → read the paper

Replacing thinking with tool usage enables reasoning in small language models
formats reasoning traces as tool interaction logs, enabling smaller models to perform complex tasks by manipulating stateful tools instead of simulating thoughts → read the paper

MindJourney: Test-Time Scaling with World Models for Spatial Reasoning
pairs a VLM with a video-diffusion world model to simulate egocentric 3D transformations at inference time, improving spatial reasoning without fine-tuning → read the paper

AgentsNet: Coordination and Collaborative Reasoning in Multi-Agent LLMs
develops a scalable benchmark to test how large-scale multi-agent LLM systems self-organize and solve distributed reasoning tasks, revealing the coordination limits of current models → read the paper

MIRIX: Multi-Agent Memory System for LLM-Based Agents
builds a modular memory architecture with six memory types and dynamic agent coordination, enabling persistent and multimodal memory for AI agents at scale → read the paper

🌟 Towards Agentic RAG with Deep Reasoning: A Survey of RAG-Reasoning Systems in LLMs
integrates retrieval and reasoning in agentic systems, outlining synergistic methods that iterate between search and inference to enhance factuality and reasoning depth → read the paper

A Survey of Context Engineering for Large Language Models
establishes context engineering as a discipline, detailing retrieval, memory, and agentic integration strategies to improve information management and reasoning in LLMs → read the paper

🌟 FlexOlmo: Open Language Models for Flexible Data Use
enables modular inference using independently trained mixture-of-experts, allowing users to include or exclude data sources at inference time without retraining → read the paper

First Return, Entropy-Eliciting Explore
stabilizes RL training for LLMs by identifying high-uncertainty steps and guiding exploration via structured semantic rollouts → read the paper

Perception-Aware Policy Optimization for Multimodal Reasoning
adds internal perception loss to reward learning, significantly reducing vision-related reasoning errors in multimodal benchmarks → read the paper

RLEP: Reinforcement Learning with Experience Replay for LLM Reasoning
enhances sample efficiency by replaying verified successful reasoning traces, improving convergence and final accuracy on math benchmarks → read the paper

One Token to Fool LLM-as-a-Judge
demonstrates vulnerabilities in generative reward models used in RL and proposes a data augmentation strategy to improve robustness → read the paper

A Survey on Latent Reasoning
explores non-verbal, internal inference mechanisms in LLMs, including hidden state propagation and infinite-depth diffusion reasoning → read the paper

🌟 Mixture-of-Recursions: Learning Dynamic Recursive Depths for Adaptive Token-Level Computation
reuses a shared stack of transformer layers across recursive depth levels to enable efficient, token-specific reasoning → read the paper

Differential Mamba
applies differential design to state-space models, improving long-range context handling and reducing hallucinations in Mamba architectures → read the paper

Scaling Laws for Optimal Data Mixtures
predicts ideal domain weightings for foundation model training using a principled scaling law framework, avoiding costly trial-and-error → read the paper

SingLoRA: Low Rank Adaptation Using a Single Matrix
stabilizes LoRA training by collapsing the adaptation into a single symmetric matrix, reducing parameter count and improving performance → read the paper

T-LoRA: Single Image Diffusion Model Customization Without Overfitting
uses timestep-aware low-rank adaptation to enable robust personalization from a single concept image in diffusion models → read the paper

🌟 Lizard: An Efficient Linearization Framework for Large Language Models
linearizes transformers using gated attention and meta-memory to support infinite context generation with constant memory → read the paper

Open Vision Reasoner: Transferring Linguistic Cognitive Behavior for Visual Reasoning
transfers cognitive behaviors learned in language tasks to visual reasoning via a two-stage cold-start and RL process, achieving state-of-the-art performance → read the paper

NeuralOS: Towards Simulating Operating Systems via Neural Generative Models
simulates GUI-based OS interactions by combining RNN state tracking with diffusion-based screen rendering, enabling AI-driven interface modeling → read the paper