I'm convinced I've been studying wrong my entire life. 🤯 A free Chrome extension just showed me how to turn ANY YouTube video into a personal AI tutor. It generates instant summaries of hour-long lectures and lets you ask complex questions without ever hitting pause. It’s the

AI in 2023: “haha it can write poems.” AI in 2025: GPT-5.2 builds spreadsheets & decks, ships real code, aces competition-level math, and runs long-horizon agents over your tools & data. We skipped a few side quests. What part of your job still feels “safe” from AI? 🤔

Genuinely curious: In your company, who is more advanced in day-to-day AI use – - senior leadership - mid-level ICs - interns / fresh grads - or students you work with? Reply with what you’re seeing 👇

If you lead a team, this has 2 implications: 1️⃣ Your interns may have more hands-on LLM experience than your managers. 2️⃣ Your AI rollout can’t just be “top-down training.” You need to tap into the students & early-career people who are already experimenting. They’re quietly

And workers? PwC’s 2025 Global Workforce Survey (~50k workers, 48 countries): - 54% say they’ve used AI in their job in the last 12 months - But only 14% use genAI daily - Just 6% use “agentic AI” daily https://t.co/hYF3x5UVU3 BCG’s 2025 AI at Work report calls it a “silicon

Now look at students + homework: Save My Exams (June 2025): - 75% of students use AI to help with homework (just 26% in 2024) - 24% use it daily, 44% weekly - 2/3 of students don’t see it as cheating, while >50% of parents do So teens aren’t just chatting for fun. They’re

Zoom in on teens ⬇️ Pew (Dec 2025): - 64% of US teens (13–17) use AI chatbots - About 3 in 10 use them every day - Nearly all (97%) are online daily, and 4 in 10 are “almost constantly” online For a big chunk of them, “ask the chatbot” is just part of thinking.

Teens are further into AI than most offices. Not vibes – the numbers: – 64% of US teens use AI chatbots; ~30% use them daily. – 75% of students use AI for homework; 24% use it daily. – But only 14% of workers use genAI every day at work. We’re onboarding a generation that’s

7/ Blunt prediction: In 12–24 months, the highest-leverage devs won’t be the ones who type fastest. They’ll be the ones who know how to: – structure Slack convos so agents can act – design prompt templates for bugs/incidents – review AI PRs faster than others can write them.

6/ But there’s a sharp edge: – Security: agent sees code + Slack history. – Misuse: sloppy prompts → overconfident fixes. – Governance: which repos, which channels, what logs? This is real DevSecOps, not just “install a bot and pray”.

5/ For eng leaders, this suggests a new operating model: • Route small issues through a #with-claude channel. • Standardize what goes in the prompt: logs, repro steps, risk. • Let Claude draft code, humans review + merge. Measure: lead time, on-call fatigue, review load.

4/ That unlocks new workflows: – Bug thread → proposed fix PR. – Feature chat → spec + scaffolded implementation. – Incident channel → postmortem + regression tests. Less “we’ll ticket this later”, more “this thread already shipped something”.

3/ Claude Code flips it: Slack = where work is *defined*. Repos = where work is *applied*. You talk about a problem → tag @Claude → it pulls context, edits code, runs tests, and drops a PR link in the same thread. Threads become “live tickets” that can end in PRs.

2/ IDE assistants were a half-step. They help individual devs, but they *don’t see the conversation*: – bug reports – customer escalations – product debates – incident channels The real context lives in Slack, not in ` https://t.co/xddiT0wV2i`.

1/ Hot take: AI "inside Slack" is a bigger shift for dev teams than AI "inside the IDE". Anthropic just launched Claude Code for Slack — an agent that reads your threads, touches your repos, and posts PRs back into chat. This changes how engineering orgs actually work.

And for builders: - Agent 1 / Copilot drafts the patch, spec, or plan - Agent 2 only tries to break it (tests, red-teaming, edge cases) - You only trust what survives both Hot take: “generator + verifier” will matter more than which logo you use. What’s the most obvious place

Mini version you can use for learning: 1️⃣ Ask AI to explain a topic 2️⃣ Then: “Now be my discriminator. Write 10 questions that should expose holes in my understanding.” 3️⃣ Answer first, then let it grade + explain If you try this, reply with the topic you tested it on — I’m

This is basically GAN energy but for thinking: - Generator: “here’s my proof / plan” - Verifier (like a discriminator): “this doesn’t hold up, try again” GANs did this for fake images vs real ones. Now we’re doing it for multi-step reasoning. Do you buy this “GANs for

Context in 2 lines: - DeepMind & OpenAI already hit IMO gold with closed models - DeepSeek Math V2 hits similar levels and ships open weights + a “generator → verifier” loop Repo if you want to peek under the hood: https://t.co/kWAyWn7KvZ But the scores aren’t the interesting

“AI got an IMO gold medal” used to melt my brain. Now it’s background noise. What’s interesting now is how these models think: they write a solution, then a second brain tries to kill it before it ships. DeepSeek’s new open math model does exactly this. Here’s why that pattern

https://t.co/DWMQ7TYWWz has made consuming youtube videos a seem-less experience. Before LearnLens : I watch youtube video. Pause it. Copy youtube url, paste to perplexity and ask my question. Come back to watching youtube video. After LearnLens : Now I simply pause Youtube