Full-text search across article titles and summaries. Refine with sector, year, content type, and reading-time filters.
Sponsored
Granola [13] listens to your calls and uses AI to add context to your shorthand notes. Then it follows up for you and produces briefs and notes for your time. No awkward meeting bot required: Granola works locally on your device. Now TLDR readers get 1 month free with code
On-policy distillation trains a student model on trajectories sampled from its own policy while a teacher provides dense token-level supervision through KL-based regularization, closing the train-inference distribution mismatch that off-policy methods suffer. The canonical formulation unifies forward-KL, reverse-KL, and JSD losses with reverse-KL emerging as the default for mode-seeking smaller students, and a one-line code swap of the regularizer model on top of an RL stack like Tinker implements the technique.
Models.dev consolidates specifications and pricing of various models, accessible via an API.
BenchBench is a benchmark that tests how well models can create a benchmark. It works as a great benchmark for model abilities as well as a test of models' self-awareness. The benchmark tests creativity and not just problem-solving ability. In tests, GPT 5.2 was the only winner, with every other model, from Opus 4.6 to GPT 5.5, struggling to create an actually useful benchmark that others had a hard time solving.
Google DeepMind's AlphaProof Nexus autonomously solved nine out of 353 open Erdős problems, including questions unanswered for decades, at inference costs of a few hundred dollars per problem.
Sponsored
Scribe Optimize [18] automatically detects how work actually gets done across your org and surfaces inefficiencies — so your AI investments are built on ground truth, not gut feel. See how work
GPT-5.6 seems heavily focused on stronger multi-step reasoning, better agentic workflows, and improved frontend generation capabilities.
Prediction markets have failed to deliver Robin Hanson's 1990 Idea Futures vision.
TLDR's Applied AI team is tasked with making every process at TLDR legible to code, runnable by anyone, and composable into larger workflows. Join a small, fast moving team using the latest AI tools with an unlimited token budget. Learn more [22].
DeepSeek's aim is to enable a $10 trillion Chinese AI hardware ecosystem and achieve a $1 trillion valuation for itself.
Apple plans to upgrade AI image tools, Genmoji and Image Playground, in iOS 27, enhancing visual quality and realism.
Sponsored
...but they're still spending on AI solution as if they do.
Anthropic appears to be moving Claude Mythos to broader availability, with the model now helping protect a wider range of organizations. Traces of the model have already surfaced on Google Cloud and AWS through vulnerability discovery programs. A release for Mythos 1 seems imminent. Claude Opus 4.8 is also rumored to be in the works for release.
DeepSeek has permanently cut V4 Pro prices by 75%. The promotion was originally scheduled to expire at the end of the month. DeepSeek's pricing sits below OpenAI's GPT-5, Anthropic's Claude Opus 4.7, and Google's Gemini 3.5 Flash. The gap is widest against the frontier reasoning models that enterprise customers rely on for demanding workloads.
Mythos Preview is capable of turning vulnerabilities into exploit primitives and combining those primitives together into complete end-to-end attack chains. The knowledge and expertise required to develop exploits will drop significantly as the model's capabilities become more widely available. This article looks at the model's performance on ExploitBench and ExploitGym, two newer, more challenging academic benchmarks. Mythos Preview consistently outperforms all other evaluated models on these tests.
Qualitative analysis is the process of reading a lot of messy, unstructured data and figuring out what is interesting, recurring, surprising, or important. One of the biggest research questions at the moment is, 'What is the right way to do agent-assisted qualitative analysis?' This post looks at the background of the field, proposes an experimental setup, and discusses how the field should move forward. It is still unclear whether AI agents can replace humans in qualitative analysis, but they can currently do the mechanical parts of qualitative analysis fast - just without taste.
SpaceX bought xAI, took the two Colossus data centers Elon built, and started renting them to Anthropic for $15 billion a year. SpaceX did $18.7 billion in total revenue last year after 23 years of building rockets, so they almost doubled the company in a quarter by becoming a cloud provider. The bigger math is wilder, the AI buildout adds up to $7.5 trillion of spend over the next 4.5 years, about 5% of US GDP and on par with the 1880s railroad boom. Even if neoclouds only catch 20% of the $13.5 trillion of value that buildout creates, that's $2.5 trillion up for grabs because no handful of hyperscalers can build it all.
Sponsored
Use these 15 agentic workflows in Notion [13] to automate, connect, and repeat processes. Use cases span ops, prod, support, and recruiting - including real examples used by Ramp, Vercel, and Clay. The playbook is free, detailed, and full of steps you can implement
The release candidate for the next Model Context Protocol (MCP) specification is now available. It is the largest revision of the protocol since launch. The candidate introduces a stateless core that scales on ordinary HTTP infrastructure, extensions, authorization that aligns more closely with OAuth and OpenID Connect deployments, a formal deprecation policy, and many other changes. It contains breaking changes. The final specification will ship on July 28.
OpenAI outlined a macro-evaluation workflow for agentic systems that analyzes patterns across entire populations of traces rather than isolated failures.
Reasonix is a DeepSeek-native coding agent for the terminal. It is engineered around prefix-cache stability and designed to be left running. Token costs stay low across long sessions.
Lance is a lightweight native unified multimodal model that supports image and video understanding, generation, and editing. It delivers strong performance across image generation, image editing, and video generation benchmarks with just 3B active parameters. The model was trained entirely from scratch within a 128-A100-GPU budget. Examples of clips generated by the model are available in the repository.
David Sacks, a venture Capitalist, warned President Trump on a call that the long-awaited executive order on the dangers posed by artificial intelligence that Trump was deliberating on could lead to mandatory regulations that slow down the industry in its race with Chinese competitors. Trump responded that he shared concerns about China and was worried about hindering AI investment. He then postponed the signing and told reporters he wouldn't sign the order. The incident shows how powerful Sacks' influence is and marks a win for those against strong guardrails to limit the risks posed by the technology.
Anthropic is on track to do $10.9 billion in Q2 revenue, up from $4.8 billion in Q1, growing faster right now than Zoom did at the peak of the pandemic. The thing that flipped them to profit is compute getting cheaper, 71 cents of compute per revenue dollar in Q1, 56 cents in Q2. Claude Code on its own is at $2.5 billion in revenue, and the company expects to clear $559 million in profit just in time for its October IPO. The "AI labs burn money forever" story finally has a hole in it.
Sponsored
clerk init scaffolds auth into your project. clerk config manages settings in code. clerk api fetches users, orgs, and sessions — no dashboard required. Open source and available for your agentic harness today.
ChatGPT users can now upload a form image, add the details they want included, and the chatbot can fill it out for them.
Perplexity open-sourced Bumblebee, a read-only security scanner that identifies risky packages, extensions, and AI tool configurations on developer machines.
TLDR's Applied AI team is tasked with making every process at TLDR legible to code, runnable by anyone, and composable into larger workflows. Join a small, fast moving team using the latest AI tools with an unlimited token budget. Learn more [24].
Gemini 3.5 Flash (Low) generates around 45% fewer tokens than Gemini 3.5 Flash (Medium) and generally outperforms Gemini 3.5 Flash (High) on SWE tasks.
Memory Files distributes notes across multiple structured documents organized by topic, project, or context.
Sponsored
Can your architecture defend against attacks that are autonomous, adaptive, and faster than anything you've seen before?
Cursor's annualized revenue hit $3 billion in late April. The company now has more than 3,000 customers paying at least $100,000 each for its software on an annualized basis. The business has become one of the fastest-growing startups of all time. SpaceX has the right to buy Cursor for $60 billion during a 30-day window that starts soon after it begins trading publicly. SpaceX is expected to list its shares on June 12.
Manus has been told to unwind its acquisition by Meta by China's regulators. Its founders now need to raise funds to buy back the operation. They are now in discussions for funding and may chip in with their own money to finance the transaction. A reversal of a deal this large months after completion is virtually unheard of.
Microsoft plans to supply its Maia AI chips to Anthropic, which currently faces compute challenges despite partnerships with Amazon and Google. Anthropic's increasing AI-assisted programming needs could benefit from Maia's 30% improved performance. A $5 billion investment by Microsoft in November strengthens this potential collaboration.
Cursor described lessons from building cloud agents, emphasizing durable execution, isolated development environments, self-healing infrastructure, and cleaner separation between agent state and conversation state.
Local, open-weight models running on older commodity hardware are becoming more and more competitive with models on the frontier. This has big implications for what frontier labs will be able to charge for their biggest models. Many applications don't need the best models, so it's not worth paying the premium.
The current rate of scaling compute isn't sustainable unless AI starts to dramatically accelerate economic growth. There is no guarantee that this rate of AI capex spending can continue after 2026. However, flat compute capex is still growing the compute stock for years. AI chips will still improve, and companies can research and train models even with a fixed amount of compute.
Sponsored
It's 10 o'clock. Do you know where your LLMs are? In this Datadog guide [13], you'll learn how to monitor LLM workflows end-to-end, detect security risks, and deploy mitigation steps. You don't need to be on high alert if your system is reliable. Read the guide to see how
Alibaba's Qwen team released Qwen3.7-Max as a proprietary agent-foundation model that posts top scores on Terminal-Bench 2.0-Terminus, SWE-Pro, SciCode, MCP-Mark, GPQA Diamond, HMMT Feb 2026, and IMOAnswerBench, with consistent performance across Claude Code, OpenClaw, Qwen Code, and custom harnesses.
Sparse autoencoders (SAEs) can help explore neural geometry by representing curved manifolds using several distinct methods, including shattering, compact capture, and dilution. The study reveals that each SAE feature only represents part of the whole manifold, necessitating the clustering of features to understand the overall structure. An unsupervised pipeline leveraging these findings could enhance our understanding of neural networks by reconstructing the internal geometric structures on their terms.
Google is aggressively integrating AI into its products to remain competitive while protecting its lucrative core businesses. The company leverages its scale, distribution, and financial resources to deploy AI like Gemini 3.5 Flash and features such as YouTube's "Ask YouTube." Despite challenges, Google aims to maintain market leadership by adapting its offerings swiftly without undermining its existing revenue streams.
OpenAI reported $5.7 billion in Q1 revenue, surpassing Anthropic. The high demand for AI inference chips has companies like Amazon, Google, and potentially Microsoft vying to supply custom solutions. Anthropic's rapid adoption by enterprises has exposed compute constraints, leading it to strike deals with multiple providers to secure capacity.
Sponsored
AI is stalling and data is the bottleneck. This Algolia white paper explains how orgs can get clean, searchable data without writing new