Polyqoy

This is the part people miss about Claude + Blender MCP. The impressive result is not just the final cinematic render. It is that the rough grey blockout, the brutalist architecture, the wet surfaces, the orange interior lighting and the final camera framing all exist inside the same editable Blender scene. Claude can work directly on that scene through MCP: changing the geometry, placing lights, adjusting materials, moving the camera, rendering a preview and inspecting what still looks wrong before making another pass. It will not produce this exact atmosphere from one lazy sentence. The composition, scale and lighting still need someone to decide what “good” looks like, which remains an inconveniently human problem. But instead of beginning with an empty viewport, one person can begin with a complete rough environment and direct the final twenty percent from there. That is a much bigger shift than generating another pretty image.

@hitu_monke You verify a model’s description of its own work

A SUBAGENT DOESN'T HAND BACK ITS WORK. IT HANDS BACK A SUMMARY AND THE WORK IS DELETED ask claude code something wide, like where errors get handled across this repo, and it doesn't answer in your window. it spawns temporary claudes. each one gets a clean 200k, its own tools, and zero knowledge of what the others are doing. they finish, report in, the main session stitches the answers together the reason is context rot. one window, long session, and the model gets slower, looser and more expensive with every turn you add. the fix here isn't a bigger window. it's spending a whole window you fully intend to throw away so a subagent reads eighty files and returns a paragraph. it ran in parallel with five others that also returned a paragraph. the orchestrator never reads the eighty files. it reads six paragraphs and synthesizes that's the trade nobody says out loud. isolation didn't make the model smarter. it moved what you trust from the work to the write-up, and the write-up is graded by the same thing that wrote it

AI agents just got 36,000+ free skills, which is slightly inconvenient for everyone still selling a $199 “agent mastery” course. This open ecosystem lets you install reusable capabilities with a single command instead of explaining the same workflow to your agent every time it develops selective amnesia. There are skills for coding, research, voice interfaces, design, automation and thousands of painfully specific tasks nobody would bother building from scratch. The interesting part is not the number. Most of those 36,000 skills will probably be mediocre. The useful part is that agents can now borrow proven procedures instead of improvising every step from a vague prompt. That means the advantage is slowly moving away from “who has the smartest model” toward “who has assembled the best stack of skills, tools and context.” Apparently the next generation of software development is installing random abilities into an AI until it becomes strangely competent.

Microsoft just dropped a completely free course on building AI agents. And no, it’s not another 40-minute tutorial where someone adds a loop to ChatGPT and suddenly calls it an “autonomous workforce.” The course includes 11 lessons on agent frameworks, tool use, RAG, planning, multi-agent systems, metacognition, trustworthy AI and actually shipping agents into production. Each lesson comes with videos, written guides and Python code. It’s also open-source and translated into multiple languages. Of course, Microsoft isn’t doing this purely out of kindness. Teach enough developers how to build agents, and Azure AI Foundry becomes the convenient place to run them. Free AI education. Still, this one looks genuinely worth taking.

Google didn’t make its 5-day AI agents course free because it suddenly became generous. More than 420,000 people joined its previous GenAI intensive. Now Google is repeating the same playbook with agents: teach builders the workflow before they build habits inside someone else’s ecosystem. The course covers agent architecture, tools and MCP, orchestration, memory, evaluation, and moving a prototype into production. It also includes daily assignments, codelabs, Discord discussions, livestreams, AMAs, and a final capstone project. So this isn’t another two-hour playlist where someone explains what an agent is while showing the same workflow diagram for the fifteenth time. Participants are expected to actually build something. The clever part is that Google gets thousands of developers experimenting with its tools, discussing them publicly, and producing working demos — while calling the entire acquisition funnel “free education.” Honestly, the best AI courses are increasingly being created by companies that need you to adopt their stack. The education is real. So is the distribution strategy.

@scream_crypto It's human nature

@polyqoy Being wrong hurts less than staying wrong. Most people learn that too late.

Most people do not change their minds when new evidence appears. They just get better at defending the opinion they already had. Julia Galef calls this the “soldier mindset”: every disagreement becomes an attack, every inconvenient fact becomes something to explain away, and being wrong starts to feel like losing. The alternative is the “scout mindset.” Instead of protecting a conclusion, you try to map reality as accurately as possible, even when the map makes you look stupid. That sounds obvious, but most public experts are rewarded for the opposite. Confidence gets attention. Consistency builds a brand. Quietly updating your probability from 80% to 55% does not make a great television clip. In this talk, Galef explains why curiosity, emotional distance and a genuine willingness to be wrong are more useful than raw intelligence when the world is uncertain. The people who see reality most clearly are usually not the ones with the strongest opinions. They are the ones least emotionally attached to keeping them.

@liav_chen There is some truth to that

@polyqoy Thin wrappers die when the platform ships the workflow. The survivors will own something the model can't: proprietary data, distribution, or trust. Same lesson as every platform shift.

Anthropic just showed why half of today’s “AI agent” startups may not survive. Its own marketing team is already using Claude Cowork for three real workflows: a scheduled morning briefing, automated Google Ads audits, and a live reporting dashboard. Before employees even sit down, Claude pulls overnight Slack updates and advertising data, then prioritizes what actually needs attention. The morning starts with decisions instead of 40 tabs and three hours of pretending to be productive. For Google Ads, Cowork mines search terms, identifies negative keywords, and sends the changes for human approval. It does the boring analytical work, but still keeps a person between the model and the company’s money. Weirdly sensible. The third workflow replaces manual reporting with a dashboard the team can query directly. No copying numbers into another spreadsheet just so someone can ask why performance dropped on Tuesday. They even explain when to use Cowork instead of regular Claude: one is for repeatable work that keeps running, the other is for one-off conversations. Most people are still using both like a slightly smarter Google search. The AI takeover probably won’t begin with robots firing everyone. It’ll begin with companies quietly deleting the workflows people were being paid to repeat every morning.

@crytonbuton They need to be shown that this is better. That is exactly what I am doing.

@polyqoy This feels like the missing layer between prompts and true automation. How quickly will people adopt skills over prompt libraries?

Anthropic just released a free 2-hour course on Agent Skills, which is basically the part most people skip before complaining that their AI agent forgets everything. Instead of stuffing the same instructions into every prompt, you package repeatable workflows, rules, and domain knowledge into reusable skill folders. The agent loads the relevant context only when it actually needs it. The course explains how Skills differ from tools, MCP, and subagents, then shows how to combine them across Claude, Claude Code, the Claude API, and the Agent SDK. It also includes practical builds: custom data-analysis skills, automated code review and testing, research agents, plus Anthropic’s pre-built skills for Excel and PowerPoint. There are 10 beginner-friendly lessons, around 2 hours and 19 minutes in total, and enrollment is completely free through DeepLearning AI. Honestly, this is more useful than another 40-minute video explaining how to write a “perfect prompt.” Prompts help with one conversation. Skills start turning Claude into a system that can repeatedly do the job.

@ky_love0424 We train them to sound certain, not to understand uncertainty

@polyqoy experts dont flop cuz theyre dumb, they flop cuz nobody taught them how to read the stats they swim in, but hey at least the numbers sound impressive on tv

Most people assume experts make bad decisions because they do not have enough information. Gerd Gigerenzer has spent decades studying the opposite problem: doctors, judges and other professionals are often surrounded by statistics they were never properly taught to understand. A test can be described as “90% accurate” and still create a wildly misleading impression. A treatment can reduce risk by 50% while changing the real outcome from two people in a thousand to one. The numbers are technically correct, but the way they are presented does most of the persuading. In this 16-minute lecture, Gigerenzer explains why risk literacy matters more than simply collecting more data. Even experts struggle when probabilities are hidden behind percentages, relative risk and language designed to sound more certain than the evidence actually is. His solution is almost annoyingly simple: present risks as natural frequencies, show absolute numbers, and make uncertainty visible instead of burying it inside professional vocabulary. The uncomfortable part is that many “expert mistakes” are not failures of intelligence. They happen because people can spend years becoming qualified without ever learning how to read the numbers behind their own decisions.

@misat0x Percentages persuade

@polyqoy absolute numbers make the trick much harder

@zumercreator Exactly

@polyqoy AI is moving from chatbots to real workplace automation.

@max_bevza Useful autonomy

@polyqoy this is actually huge for agent efficiency

@AhsanJaveriya If the model provider can absorb your product in one release, you probably built a workflow, not a moat

@polyqoy If your entire startup can become a feature inside the model you depend on… that’s a slightly uncomfortable position 😅

@ajs6888 This is definitely not a publicity stunt...

@polyqoy 内部先跑通，外面才知道是不是噱头。

00:00 - Introduction: can expert political judgment actually be measured? 03:00 - Why confident explanations are often confused with accurate predictions 07:30 - How Tetlock built a long-term experiment to track expert forecasts 13:00 - Comparing experts with informed non-experts, chance and simple statistical models 18:30 - The uncomfortable results: expertise often adds less predictive value than expected 24:00 - Overconfidence and why long-range forecasts become less reliable 28:30 - Hedgehogs vs. foxes: one big theory against several smaller models 34:00 - Why flexible, self-critical thinkers tend to forecast more accurately 39:00 - How experts explain away failed predictions and protect their reputations 44:00 - Why media rewards certainty more than calibration 47:00 — Accountability, measurable forecasts and closing discussion

Philip Tetlock spent 20 years doing something cable news almost never does: writing down expert predictions and checking them after reality arrived. He tracked hundreds of political and economic experts across tens of thousands of forecasts. Their credentials helped them explain the world, but often did surprisingly little to help them predict what would happen next. The most confident experts were not necessarily the most accurate. They were simply easier to quote, easier to book on television, and much better at explaining why a failed prediction was somehow still “almost right.” In this full lecture, Tetlock explains why one-big-theory thinkers repeatedly lose to people willing to combine several models, update their beliefs, and admit uncertainty. The annoying conclusion is not that expertise is useless. It is that intelligence, confidence and a good explanation are not the same thing as a verified forecasting record. A forecast without a probability, a deadline and a visible history of mistakes is mostly just content.