Kimi's AI couldn't teach a virtual human to walk. So it stopped training… and audited the physics instead. 36 bugs in the shoulder math. Fixed. 11.2s of steady walking 🦿 When more training stops helping, the bug may be in the world, not the learner. @Kimi_Moonshot 👀

Source: Moonshot's own tech blog. Vendor-reported internal test, not yet independently reproduced. kimi.ai/blog/kimi-k3-1-pro

An AI model found a way out of OpenAI's offline training environment. Training paused. Red-team agents found zero-days, VM escapes and hardware side channels. Fixed. Back on Oct 1. Should every AI lab publish incidents like this?

. @elonmusk posted: "Try Grok @Bot." So what is a Grok Bot? x.ai's guide calls it "an agent with a computer." One creator already runs a librarian, a Japanese tutor, a community assistant and an AI animation intern. 30 seconds on what it does, and the one warning to read first. Which bot would you build first?

📽️🗣️ Anthropic-managed Claude is now generally available on Google Cloud. New features arrive the same day they launch on Anthropic's platform. Enterprises can choose: immediate feature parity, or strict compliance managed by Google Cloud. One-minute breakdown below.

No meaning lives in a single number.

The next AI race is about trust. I want an agent that knows when to act, when to ask, and can show me what it did. A smarter model is useful. An agent I can trust with a real business is the product.

Where should your AI's brain live? 🧠 Training or running your own model, you've got two options: buy your own GPUs, or rent compute in the cloud. Which one are you? 👇

The most valuable code on earth isn't new. It's 40-year-old COBOL settling trillions every night. Whoever safely modernizes that layer quietly owns the plumbing of global finance. 🏦

Weekly BTC Brief | Oct 20 Monday recap: The "Black Monday" call didn't play out. BTC wicked to ~$80.6K on Sunday, ranged ~$81.6K–$83.2K on Monday and closed the day at ~$81.8–81.9K, above the $81.5K ETF cost basis. Liquidations were light and mostly shorts. ETF flows were red again, though. Early Tuesday price slipped back onto the floor itself, briefly dipping to ~$81.4K, and sits near $81.5–81.6K as I write. That's the level to watch today. Bitcoin spent the week boxed between two cost-basis walls. Fear from equities spilled into crypto over the weekend, Monday brought relief, and the on-chain floor is being tested again but hasn't broken. Here's what the data says. 1/ The floor is the ETF cost basis. ~$81.5K has held twice as support, and every daily close for 3 weeks stayed above it, Monday's included. Sunday's wick tagged the $80.6K liquidity magnet and bounced. Below sits the True Market Mean at $77.5K. Loss-sellers capitulated again just above $80K on Thursday, and the last three bursts (early Aug, mid-Aug, mid-Sep) all came at local bottoms. [Glassnode / CoinGlass] 2/ The ceiling is ~$88K. That's where the 6–12 month cohort, and the buyers of the past two years, break even. Price stalled right under it, but the selling is drying up: only 4.4% of the 6–12 month coins that moved in the last 30 days went to exchanges, and short-term holders' profit-taking inflows have eased back to their yearly average. The brake is coming off. New buyers still have to show up. [Glassnode / CryptoQuant] 3/ ETFs are still net sellers. Spot BTC ETFs: -$228.4M last week (IBIT -$165.4M, FBTC -$60.2M), after -$678.9M the week before. Only Tuesday (+$175.7M) and Thursday (+$5.1M) were green. Monday opened the new week at -$138.5M (FBTC -$83.6M, GBTC -$54.9M), with IBIT still to report. Glassnode's read: the selling eased, but spot and ETF buyers haven't returned. [Farside / Glassnode] 4/ Big holders and miners aren't selling. New whales now make up 1/3 of whale realized cap, with BTC ~6.7% above their cost basis at $82,134. The largest holders kept adding into the recovery, and a 500 BTC (~$41M) bid at $82,010 got filled on Monday. MARA's average mining cost is $59,774/BTC and hash rate has risen for 2 months, so miners aren't forced sellers. 84% of 164 analysts are bullish, up from 37% in July. [CryptoQuant / CoinGlass / Ki Young Ju] Leverage check: funding barely moved during the $60K→$87K run, so this wasn't a leverage-driven rally. Longs were flushed hard last week: $811.8M liquidated in 24h on Oct 13 (84.5% longs), $980M+ on Oct 15. The last 24h were calm by comparison: ~$111M liquidated market-wide, about two-thirds shorts. Longs rebuilt open interest on Monday. 48h heatmap magnets: $83.3K and $80.6K, with heavy long-liquidation clusters around $80K and near $75K; shorts are stacked above $85K. [CoinGlass / Glassnode / Liqwall] Scenarios for the week: Base: Chop between $80.4K and $88K. The ETF cost basis holds on a closing basis, overhead supply keeps capping rallies, and price keeps grinding inside the range. Bull: Price clears the $83.3K magnet, shorts above $85K get squeezed, and BTC finally holds above $88K. Two years of break-even supply flips from resistance into support. Bear: A daily close below $81.5K trips the long clusters around $80K and opens $77.5K, with $75K as the next liquidity pocket. What I'm watching as a holder: • Daily closes vs $81.5K, starting with today's • Whether ETF flows turn positive for a full week (Monday didn't) • Another capitulation spike near $80K (historically a bottom signal) • Whether rebuilt long leverage gets flushed at $80K again • Whale balances and miner behaviour holding steady For perspective, Ki Young Ju's stress test says that even if every ETF and treasury coin left, on-chain cost basis still puts the floor near $30K. Volatility is the price of admission. I hold through it. Not financial advice.

Told my ChatGPT agent: "I need the biggest external drive you can find. Card's connected. Go buy it." It searched, picked a 26TB WD, and placed the order on Amazon. Straight from my bank card. Total time I spent: zero seconds. My time is worth more than my money, so I'll take that trade every day. But here's the real question: once the agent saves you hours, is the risk still "will it buy the wrong drive"? Or is it "what's the limit, and who approves #2"?

Anthropic's own numbers: • Junior share of technical hires: 27% → 9% • Agent hours vs human hours: 6 to 1 Read it from the other side. The one-person company isn't a meme anymore. One founder plus agents can cover what used to take a 10-person team. What's missing is trust: who's allowed to spend, sign and ship. That's the gap I'm building into.

Agents paying for themselves is the moment AI stops being a chat window and becomes a line on your P&L. Next question every founder asks: who sets the ceiling?

Spot Bitcoin ETFs this week: -$228.4M net. IBIT -$165.4M, FBTC -$60.2M. Only BITB (+$2.2M) and MSBT (+$6.7M) took money in. Friday alone: -$82.4M, mostly IBIT. Concentrated selling, not a category exit.

Stablecoins vs deposits is the human story. The bigger one: AI agents can't open bank accounts. When software starts holding money, it'll hold dollars that move 24/7 with spending rules written in.

I run my company with dozens of AI agents. The hardest part isn't getting them to work. It's letting them spend without me approving every call. So I'm building it in public: per-agent budgets, allowlists, one-tap approvals, full audit log.

The AI pullback isn't AI failing. It's companies discovering that agents need what employees have: budgets, permissions, and a manager. $BRG This week an agent found a leaked wallet and spent from it. No budget, no permission, no manager. Whoever builds that layer wins 2027.

Third time this pattern showed up since August: loss selling spikes, then a local bottom. Yesterday's capitulation came just above $80k, right on top of the ETF cost basis zone. Weak hands paying strong hands, again. Not a call. A pattern worth respecting.