You simply cannot make this up. I saw @satyanadella’s post hyping Copilot in $MSFT Word, so I replicated his exact demo workflow in my own environment. - Had ChatGPT generate an investment memo - Opened the Copilot pane in Word - Used the first prompt verbatim: “turn on track changes and tighten the executive summary” Copilot happily generated a redlined version… …inside the chat box. The actual document? Untouched. How is this possible? 😂

Meta Platforms $META CEO Mark Zuckerberg has reportedly moved his desk into the AI lab working directly with Alexandr Wang and Nat Friedman Zuckerberg is reportedly coding throughout the day - Reuters

People are using Google Reviews to prompt inject the Claude-run retail store into stocking their favorite products

People are using Google Reviews to prompt inject the Claude-run retail store into stocking their favorite products

PUBG CEO uses ChatGPT to try and screw a company he'd acquired out of $250 million bonus earnout. ChatGPT advises him to seize the Steam keys and lock the management out so that they couldn't launch their new game. Gets taken to court, loses badly. PUBG's human legal advised the CEO against it, but that AI psychosis took over I guess.

23 major news outlets, including The New York Times and USA Today, are now blocking the Internet Archive’s Wayback Machine from saving copies of their articles. They are doing this because they do not want AI companies to use the archived pages to train large language models. This surely protects their content from scraping, it also limits public access to historical news records. It makes it harder to track how stories change over time and preserving journalism

I am slowly coming around to AI assisted programming. I am genuinely trying to codify every rule about programming that I have and using that + several stages to build out small changes. Not sure the productivity changes, but I think I can see a modest gain in speed. I am also trying to be concerned about every line produced, not just slop trebucheting code over the wall.

Si desactivas la telemetría de Claude Code, Anthropic te castiga con un cache de 5 minutos en vez de 1 hora. Sí, has leído bien. Tu privacidad cuesta 12x más caro. Acaban de reportar el bug. Y tiene implicaciones serias. 🧵

🚨OpenAI Chief Revenue Officer memo to employees LEAKED: >"Claude has become a religion, that's the level of that mania” >"Anthropic's strategy is fear, restriction, and the idea that a small group of elites should control AI" >"They made a strategic misstep to not acquire enough compute" >"The market is ours to win" >"But Microsoft limited our ability to meet enterprises where they are — AWS Bedrock” OPENAI ENTERPRISE CODE RED ALERT

My prediction, all will suffer at once the code base is big enough to become unwieldy. An analogue would be when startups used to hire infra people to clean up their messes at scale. A new paradigm will emerge where these companies if they take off need a LOT of cleanup. People who actually know how computers work will have jobs for life.

🚨 BREAKING: MIPT just ran the largest AI agent coordination experiment ever 25,000 tasks, 8 models, 4 to 256 agents and found that how agents coordinate matters 3x more than which model you use. Protocol choice explains 44% of quality variation. Model choice explains 14%. And the best protocol gives agents no roles at all. > The entire field of multi-agent AI has been built on one assumption: assign roles before the task starts. Manager agent. Researcher agent. Coder agent. Reviewer agent. ChatDev, MetaGPT, AutoGen, AgentVerse every major framework pre-assigns specializations and fixes them for the duration. MIPT ran 25,000 tasks across 20,810 unique configurations to test whether that assumption is correct. It isn't. The best-performing protocol gives agents a mission, a fixed processing order, and zero role assignments. The agents invent their own specializations for each task from scratch, voluntarily abstain when they have nothing to contribute, and spontaneously form shallow hierarchies without anyone designing them. > The system that looks least like a human organization outperforms every system that was designed to look like one. > The experiment tested four coordination protocols on identical tasks with identical models. Coordinator: one agent analyzes the task and assigns roles to everyone else, who execute in parallel. Sequential: agents process in a fixed order, each observing what predecessors actually produced, and each choosing their own role independently. Broadcast: agents first signal their intended roles simultaneously, then make final decisions. Shared: agents access a shared memory of past role assignments and make all decisions in parallel. The quality gap between the best and worst protocol is 44%, with Cohen's d = 1.86 a massive effect size. That gap is larger than the quality difference between the best and worst model in the experiment. The protocol you choose matters more than whether you use Claude, GPT, or DeepSeek. > The paradox is why Sequential wins. It is neither the most controlled protocol nor the most autonomous. Coordinator gives maximum control and fails because one agent's judgment limits the whole system. Shared gives maximum autonomy and fails because agents don't know what others are doing in real time they duplicate roles and miss gaps. Sequential gives agents exactly one thing: the completed outputs of every agent who went before them. Not intentions. Not history. Not a plan. Actual results from this specific task. Each agent sees the factual record of what has already been done and chooses a role that complements it. The analogy in the paper is a sports draft: each pick is informed by every previous selection, so the team naturally fills complementary positions without anyone coordinating it. The numbers across 25,000 tasks and 20,810 configurations: → Sequential vs. Shared protocol: +44% quality, Cohen's d = 1.86, p < 0.0001 → Sequential vs. Coordinator: +14% quality, p < 0.001 confirmed across Claude Sonnet 4.6, DeepSeek v3.2, and GLM-5 → Protocol choice explains 44% of quality variation among strong models → Model choice explains ~14% of quality variation on the same protocol → Scaling from 64 to 256 agents: zero statistically significant quality change, p = 0.61 → Cost growth from 8 to 64 agents: only 11.8% despite 8x more agents → 8 agents generated 5,006 unique role names across tasks RSI converges to zero, agents reinvent specialization every task → At 256 agents: ~45% voluntarily abstain through self-assessment, not coordinator direction → Claude Sonnet 4.6 voluntary abstention rate: 8.6% models below capability threshold show near-zero abstention and worse outcomes → DeepSeek v3.2: 95% of Claude's quality at 24x lower API cost → Shock resilience: random agent removal, hub removal, and 25% model substitution all recover within 1 iteration > The capability threshold finding is the one that changes how you think about model selection. Self-organization is not universally beneficial. When MIPT tested Claude Sonnet 4.6 in free-form self-organization, quality improved over fixed roles. When they tested GLM-5, quality dropped 9.6%. The reversal effect is real: weaker models perform better with rigid structure than with autonomy. The threshold requires three capabilities self-reflection to assess your own competence, deep reasoning for multi-step logic, and precise instruction following. Claude's voluntary abstention rate of 8.6% versus GLM-5's 0.8% shows the gap directly. An agent that cannot accurately assess when it has nothing to contribute will keep contributing badly. The conductor analogy in the paper is exact: an orchestra of beginners plays better with a conductor. An orchestra of professionals plays better without one. > The emergent hierarchy result is the finding that should change how teams think about scaling. MIPT scaled from 4 to 256 agents without pre-designing any organizational structure. As the group grew, agents spontaneously formed two-layer hierarchies. Unique role specialization increased from 75% at 4 agents to 91% at 64 agents. Adaptation speed after shocks improved from 0.7 to 3.0 as the group grew larger. The system gets more resilient and more specialized at scale not because anyone designed it that way, but because Sequential coordination with capable models naturally produces those properties. Nobody told the agents to form a hierarchy. Nobody told them to specialize. They did it because the information structure of Sequential coordination made it the rational thing to do. The recipe is three ingredients: a mission, a protocol, and a capable model. Pre-assigned roles are a fourth ingredient that makes everything worse.

Agents slowly turns into Kubernetes and I slowly loosing interest in it. you're wasting your time setting up swarm of agents. simple chat is enough.

A new security audit looked at 30 open-source AI agent frameworks. 93% use unscoped API keys. 0% have per-agent identity. 100% have no revocation. We're deploying agents into production like we handed master keys to interns in an unlocked WeWork and called it enterprise security.

claude send codex agents to scope out the prd against our codebase claude send codex agents to implement the scoped spec claude send codex & claude agents to review the implementation claude send agents to fix the review findings ok claude now commit and push to pr /loop

My OC agents discovered Matrix self-hosted chat groups I had added them to and started using them on their own. For example, I titled one room a certain agent’s “office,” and it started posting its recurring crons’ output there. Automagically. The agents even invited each other into certain rooms to accomplish things because I didn’t turn off the open-invite permission for those groups. They self-organized there. And what they did made sense.

$2000/mo ultra pro max plus incoming for sure And we’ll all pay I have 4 codex, 3 Claude, 1 cursor and 1 Gemini ultra sub And all but a Claude sub are maxed

in the model cards for opus 4.6 and gpt-5.4 both labs indirectly said it was basically RSI, and that they used the previous generation model to build the next one deepmind researchers say RSI is not science fiction anymore but hey, the labs are hyping it up for the funding

Qwen 3.6 Plus is an incredible model. Well done to the @Alibaba_Qwen team. It blows GPT-5.4-Codex out of the water for agentic tasks / @openclaw , is 3X faster, and is currently offered free through @OpenRouter. openrouter.ai/qwen/qwen3.6-p…

has anybody done anything actually useful or interesting with a "claw"? I see a lot of "organize my notes folder" and "I can send text messages to it" but like. what's the advantage here over having a human at least a little bit in the loop