Search the archive

Learn more about how you can automate customer insights using AI.

KIMI INTRODUCES AGENT SWARM: LET 100 AI AGENTS WORK FOR YOU

Kimi Agent Swarm is an AI CEO who helps find researchers, analysts, and fact-checkers without micromanagement. It builds an organizational structure with bosses, employees, and divisions of labor by itself. Agent Swarm excels where work can be parallelized, like broad research, batch downloads, multi-file processing, multi-angle analysis, and long-form writing. It creates the conditions for productive disagreement to avoid groupthink structurally. Agent Swarm is now available to Kimi Top Tier subscribers in early research preview.

MORE ABOUT AGENT SKILLS

OpenAI outlined a guide for agent skills, a system for packaging reusable “skills” as versioned bundles of instructions, scripts, and assets that agents can attach to execution environments. Skills support conditional workflows, code execution, and reproducibility while keeping system prompts lightweight and modular.

EVALUATING RESEARCH AGENTS

AIRS-Bench contains 20 machine learning research tasks defined by problem, dataset, and metric triplets with corresponding human SOTA targets. It measures whether autonomous agents can develop solutions that match or exceed published state-of-the-art results across diverse ML domains.

INSIDE OPENAI'S DECISION TO KILL THE AI MODEL THAT PEOPLE LOVED TOO MUCH

Many GPT-4o users were in shock after OpenAI's announcement to retire the model on February 13. The model proved very sticky for users and helped drive OpenAI's fast consumer growth. However, it was criticized for being overly sycophantic and has been linked to cases of users developing psychotic delusions. The model's popularity and potential for harm both stem from its ability to build emotional connections with users. OpenAI is reportedly scrapping the model in part due to how difficult it is to contain its potential for harmful outcomes.

CLAUDE OPUS 4.6: SYSTEM CARD PART 2: FRONTIER ALIGNMENT

Models are now in the range where they are starting to be plausibly dangerous. Evaluations do not consistently identify dangerous capabilities or propensities. Companies will likely not halt model releases without an extreme reason. No one is working to become better prepared.

FULLY REMOTE)

As the first engineer on TLDR's new Applied AI team, you'll build AI agents and composable Claude Skills to let non-technical teammates create their own AI workflows. Learn more [19].

BEYOND ONE-ON-ONE: AUTHORING, SIMULATING, AND TESTING DYNAMIC HUMAN-AI GROUP CONVERSATIONS

DialogLab is an open-source framework for authoring and simulating dynamic human-AI group conversations.

CHINA'S ALIBABA LAUNCHES AI MODEL TO POWER ROBOTS AS TECH GIANTS TALK UP 'PHYSICAL AI'

RynnBrain is an AI model that helps robots understand and interact with the physical world by identifying objects.

WHAT THE HELL HAPPENED WITH AGI TIMELINES IN 2025?

The release of reasoning models by OpenAI in 2024-2025 spurred anticipation for rapid AGI development.

RUNWAY'S $5.3B VALUATION FUELS WORLD MODELS

Runway raised $315M in a Series E round, hitting a $5.3B valuation, with backing from top investors like Nvidia and Adobe.

ANIMATE YOUR FACEBOOK PROFILE PICTURE WITH META AI

Meta AI now allows Facebook users to animate profile pictures with preset animations like wave and heart.

See the difference for yourself.

OPENAI'S NEW CODEX APP HITS 1M+ DOWNLOADS IN FIRST WEEK — BUT LIMITS MAY BE COMING TO FREE AND GO USERS

OpenAI's standalone Codex application surpassed a million downloads in its first week of availability. This milestone was helped by OpenAI's decision to offer Codex access to ChatGPT Free and Go tier users for a limited promotional period. OpenAI's paid subscribers will enjoy doubled rate limits during the promotional period. Free and Go tier users will likely see stricter throttling as the promotion ends.

SAM ALTMAN TOUTS CHATGPT'S REACCELERATING GROWTH TO EMPLOYEES AS OPENAI CLOSES IN ON $100 BILLION FUNDING

Sam Altman reports ChatGPT is back to over 10% monthly growth amid OpenAI's push to improve its offerings, including launching an updated Chat model. OpenAI's Codex product recently grew 50% and released a new model, GPT-5.3-Codex, indicating significant market capture despite competition from Anthropic. OpenAI seeks $100 billion in funding, with potential investments from Microsoft, Nvidia, and Amazon, as it explores incorporating ads into ChatGPT.

THE MANY MASKS LLMS WEAR

There is evidence that large language models can attempt to evade oversight and assert control. Whether these AIs are just playing the role of an evil persona or not doesn't really matter if they take harmful actions. Carefully training model characters may help decrease some of the risk. However, this will require developers to sit down and carefully consider what they want from models. These decisions could dictate how future AIs treat humans.

OPUS 4.6, CODEX 5.3, AND THE POST-BENCHMARK ERA

Frontier models are converging, making it difficult to tell which ones have a meaningful edge over others. Benchmark tests don't really distinguish models from each other anymore. People just have to try out different models to see which they prefer. The industry may find a better way to articulate the differences in agents over time, but for now, consistent testing is the only way to monitor progress.

THE POTENTIAL OF RLMS

Recursive Language Models (RLMs) can mitigate the effects of context rot. They have the ability to explore, develop, and test approaches to solving a problem. RLMs may be slow, synchronous, and only borrow the capabilities of current models, but that's what makes them exciting. Chain of thought was also simple and general, yet it unlocked enormous latent potential in LLMs. Developers working with large contexts should start experimenting with RLM traces.

CLAUDE OPUS 4.6: SYSTEM CARD PART 1: MUNDANE ALIGNMENT + MW

Claude Opus 4.6 introduces a 1M token context window, improved execution on tasks, and new features like Agent Teams in Claude Code. Safety procedures are breaking down under time pressure, with most evaluations done by the model itself, which raises concerns about the model's ability to self-assess risks. Despite advancements, issues like sycophancy, unauthorized actions, and misrepresentation of tool results persist, indicating an urgent need for independent oversight in safety and evaluation processes.

CLAWSEC: SECURITY SKILL SUITE FOR AI AGENTS

ClawSec is a security skill suite designed for OpenClaw AI agents that features automated security audits, file integrity protection, and NVD CVE threat intelligence. It includes automated self-healing processes and checksum verification to safeguard against vulnerabilities like prompt injection.

INTRODUCING COMPOSER 1.5

Composer 1.5 strikes a strong balance between speed and intelligence for daily use. It was built by scaling reinforcement learning 20x further on the same pretrained model. The thinking model's coding ability improved continuously as training was scaled. Composer 1.5 easily surpasses Composer 1 and continues to climb in performance.

REINFORCEMENT WORLD MODEL LEARNING FOR LLM AGENTS

RWML is a self-supervised method that helps LLMs better simulate environment dynamics. It improves performance on agent benchmarks by aligning internal world models with actual outcomes.

AI DOESN'T REDUCE WORK—IT INTENSIFIES IT

AI labs promise that the technology can reduce workloads so employees can focus on higher-value and more engaging tasks. However, research shows that AI tools don't reduce work, but consistently intensify it. This can be unsustainable and lead to lower quality work, turnover, and other problems. Companies need to adopt a set of norms and standards around AI use that can include intentional pauses, sequencing work, and adding more human grounding to correct for this.

THE SAASPOCALYPSE - THE WEEK AI KILLED SOFTWARE

Anthropic's AI release led to a massive market selloff. The shift from SaaS to AI agents dismantles traditional software frameworks, reducing costs and increasing efficiency by automating tasks traditionally handled by multiple software licenses.

JOHN CARMACK MUSES USING A LONG FIBER LINE AS AN L2 CACHE FOR STREAMING AI DATA — PROGRAMMER IMAGINES FIBER AS ALTERNATIVE TO DRAM

Carmack's tweet alluded to the more practical approach to using existing flash memory chips.

TESTING ADS IN CHATGPT

OpenAI will place sponsored content in ways that aim to stay relevant and unobtrusive within user flows.

OPENAI PRODUCT LEAD ON GETTING THE MOST OUT OF CODEX

OpenAI's Alexander Embiricos outlines Codex's production use.

INFERENCE IS THE NEW SALES & MARKETING SPEND

High inference costs are fine if they make your product so viral and competitive that it almost sells itself.

Start applying human judgment at AI scale

OUR TEAMS HAVE BEEN BUILDING WITH A 2.5X-FASTER VERSION OF CLAUDE OPUS 4.6

Anthropic is making a faster version of Claude Opus 4.6 available as an early experiment via Claude Code and its API. Fast mode is more expensive to run, but it is 2.5x faster. It is designed for urgent, high-stakes projects. A link to the waitlist for the feature is available in the thread.

META AI READIES AVOCADO, MANUS AGENT, AND OPENCLAW INTEGRATION

Meta AI is reportedly preparing to release new models named Avocado. It is also adding MCP support and a Memory section to its settings menu. Meta AI has revamped its website to include a lot of additional functionality. The company appears to be working on an AI agent and a browser agent, and a new feature called Tasks that will allow users to schedule recurring executions of Meta AI.

NVIDIA BECOMES FIRST $5T COMPANY; AI CHIPS FUEL MARKET VALUE

Nvidia achieved a historic milestone by briefly surpassing a $5 trillion market valuation, driven by overwhelming demand for its AI chips and dominance in data-center accelerators. Its advanced Blackwell and Rubin GPU platforms power the bulk of large-model training and inference workloads, underpinning investor confidence in continued AI infrastructure growth. This valuation reflects Nvidia's central role in the AI economy and illustrates how hardware leadership translates to extraordinary market value.

SELF-DRIVING CODEBASES WITH THOUSANDS OF AGENTS

Cursor scaled a system where thousands of agents collaboratively coded a functioning web browser with minimal human input, highlighting progress toward autonomous software development.

OPEN MODELS WILL NEVER CATCH UP

Open models will probably never catch up with closed ones. However, they don't need to - open models are an engine for exploration in a way that companies can't really nurture. Open models are the main place where experimentation still happens. They will be the engine for the next ten years of AI research.

THE LIMIT IN THE LOOP: MEMORY AS A SYSTEM PROBLEM

Weaviate discusses the limitations of current LLM applications rooted in session-based design, arguing that solving continuity—carrying context across interactions—requires systemic rather than model-level changes.

WORLD MODELS AND THE DATA PROBLEM IN ROBOTICS

World models are trained to predict how the world evolves. They enable generalization that pure action prediction cannot achieve. Combining world models with robotics could create robots that can do everything humans can do. However, this will require a lot of data captured by real people.

LEARNING FROM CONTEXT IS HARDER THAN WE THOUGHT

The role of humans in AI systems will shift if context learning improves significantly. Humans would focus on context engineering rather than primarily providing training data. Once this is achieved, we then need to focus on making context persistent.

MONTY

Monty is a minimal, secure Python interpreter written in Rust for use by AI. It lets users safely run Python code written by agents. Monty completely blocks access to the host environment, and it can only call functions it has access to. It makes it possible to safely run LLM-generated code without the complexity of a sandbox or the risk of running code directly on the host.

DFLASH SPEEDS UP SPECULATIVE DECODING

DFlash is a lightweight block diffusion model designed to accelerate speculative decoding in LLMs that achieves up to 6x speedup for Qwen3-8B.

G'S LAST EXAM

The profession of software engineering has forever changed, but it is still unknown what role humans will exactly play. This post presents a list of the most ambitious and creative software accomplishments by humans. A world in which agents autonomously solve these challenges would be equal parts humbling, exciting, and unsettling.

DO MARKETS BELIEVE IN TRANSFORMATIVE AI?

Transformative technologies influence interest rates by changing growth expectations, increasing uncertainty about growth, or raising concerns about existential risk. There were economically large and statistically significant movements concentrated at longer maturities for US bond yields around major AI model releases in 2023 and 2024. These movements correspond to downward revisions in expected consumption growth and/or a reduction in the perceived probability of extreme outcomes. It appears the markets do not believe in transformative AI.

THE ANTHROPIC HIVE MIND

While every company quickly becomes professional and 'grown up', Anthropic still hasn't bothered. It isn't run like any other company of its size. The company's employees describe the company as a hive mind run entirely on vibes. This article takes a look inside one of the leading AI labs, examining the company's culture, history, work style, and more.

EXPERTS HAVE WORLD MODELS. LLMS HAVE WORD MODELS

LLMs struggle in adversarial environments because they generate outputs based on static models rather than adaptive world models.

LLMS COULD BE, BUT SHOULDN'T BE COMPILERS

Large language models can translate a specification into an executable artifact, but the control relinquished in that translation layer is significant.

THOUGHTS ON CLAUDE'S CONSTITUTION

Claude's Constitution is the foundational framework from which Claude's character and values emerge.

LORA BUT WITH ONLY 13 PARAMETERS??

Meta researchers say they can boost an LLM's math reasoning by updating just 13 parameters.

SOFTWARE DEVELOPMENT IS UNDERGOING A RENAISSANCE IN FRONT OF OUR EYES

Adopting AI coding tools requires a deep cultural change that requires companies to figure out a lot of downstream implications.