words of wisdom... since 1923 #hodl #btc #investing

The Department of Government Efficiency closed its doors on July 4, 2026, and Dogecoin is trading near $0.096 as if nobody sent it the memo. The Musk-adjacent halo that carried the coin through 2025 has formally expired, yet DOGE still posted a 20 to 24% gain over the past 30 days before pulling back about 16% from $0.10. A dead narrative can coexist with a rally when the float is in few enough hands. Wallets holding 100M+ DOGE now control a record 108.52B tokens, roughly $11.8B, while AmbCrypto describes futures participation as losing steam. Large holders plus ETF wrappers like 21Shares' TDOG and Grayscale's GDOG can move price on modest volume because retail is no longer the marginal buyer, so the chart looks healthier than the order book. Concentration cuts both ways. A market that lifts on 400M DOGE of whale buying in five days can drop just as fast when one big holder rotates, which is what Remixpoint did on September 1, selling 2.8M DOGE at a loss and keeping only bitcoin. Ali Martinez, who tracked a $15 target for years, has conceded that call is invalidated. The levels that matter now are $0.0813 support below and $0.1552 above, with futures open interest as the tell for whether anyone besides the whales is showing up.

Whale accumulation has been the bull case in every crypto cycle, mostly because nobody files the story when whales sell. Santiment counted 113,950 BTC (about $9.6B) added by wallets holding 100 to 1,000 BTC since July 15, published September 25 as Bitcoin tapped $87,000. The number is real. The label is doing the heavy lifting. A wallet in that band could be an ETF authorized participant staging inventory, an exchange consolidating cold storage, a market maker hedging a book, or a treasury firm buying on a schedule. All four print the same on-chain signature as a conviction buyer. Only one of them deserves the "smart money" tag, and the address cannot tell you which. Then comes the selection step. The 20,000-wallet Ethereum dataset that produced the UNI, LINK and ONDO inflow headlines also has a distribution column. That column rarely makes it into print, so readers see whales buying roughly 100% of the time. XRP shows the ceiling on the trade. Whales added 300M tokens in 96 hours during the run from under $1.00 to $1.70, then Fed Chair Kevin Warsh's Jackson Hole speech took about 20% back off the top. Large holders can absorb supply. They do not set the rate path. Before acting on the next accumulation headline, check which cohort is buying, who is selling into them, and whether the same dataset shows outflows the article skipped.

Galaxy Digital moved $100 million of its own balance sheet into sUSDS, Sky Protocol's yield-bearing stablecoin, and approved the same asset as eligible collateral across an institutional loan book averaging $1.4 billion. An undisclosed SKY purchase came bundled in, with neither side publishing size or price. The mechanism at work is one every bond desk already runs. Post a Treasury as security and the coupon keeps landing while the loan is outstanding. Onchain borrowers used to face a worse menu: post the stablecoin and earn zero, or hold it for yield and lose the borrowing capacity. Galaxy's 1,600-plus counterparties can now post sUSDS, keep collecting the Sky Savings Rate, and borrow against it, a repo book in all but name. The plumbing predates the headline. Grove, inside Sky's orbit, already runs a $500 million warehouse facility financing Galaxy's digital-asset-backed loans, and Galaxy has borrowed through Spark. Sky used to sit outside as a supplier of lending capacity; now one of its assets sits on Galaxy's treasury and inside its credit book as accepted security, which blurs partner into exposure. The skeptics keep a seat at the table. sUSDS is a claim on a DeFi protocol, carrying smart contract and governance risk, and its liquidity has never been tested in a proper stress event. Watch for a second institutional desk to accept sUSDS as collateral, because one adopter is an experiment and two is a standard.

Nene Royal & Ozone Band will kick off a new generation of metal/rock/grunge cycle... Such a talent! #AGT2026 #SoundGarden src: IG ozonebandfc

Michael Burry has a new number: S&P 500 net investment hit 2.07% of GDP in Q2, the highest reading in almost 40 years outside the 2000 peak itself. He expects it to pass the dot-com level within a few quarters. The word doing the work there is "net". Net investment is gross capex minus depreciation. Depreciation lags spending by design, so a buildout looks affordable right up until the bills for last year's hardware arrive. After 2000, the depreciation from all that fibre and equipment grew large enough to hold net investment negative for 12 straight quarters between 2003 and 2006. That hangover is part of what took down Lehman, Bear Stearns, Washington Mutual, Countrywide and AIG. Timing is the other lesson. In 2000 capex kept climbing after the Nasdaq had already topped. In 2007 net investment peaked months before the S&P 500 did. In the shale boom energy stocks rolled over in the middle of the drilling cycle. Spending peaks have been a poor guide to price peaks in every one of those episodes. This time round the top five hyperscalers carry $3 trillion in combined purchase commitments, leases and debt guarantees tied to AI data centres, and their free cash flow is turning negative as borrowing accelerates to fund it. Every GPU bought today becomes a depreciation charge for years ahead, and that charge lands on the income statement whether or not the AI revenue does. Watch the free cash flow line at those five companies before watching the capex headlines.

Singapore Foreign Minister Vivian Balakrishnan used his Sept 26 speech at the UN General Assembly to propose a UN Framework Convention on AI Safeguards, then admitted in the same breath that a safety pause is dead on arrival. The reason he gave was money. Superpower rivalry plus the financial incentives already locked into companies that own frontier models means nobody with a model will volunteer to park it. In his own car analogy, the engine is bought and paid for, so the only negotiable item left is the brakes. A framework convention, modelled on the UN climate treaty, bans nothing on day one. It writes down principles, stands up an institution, and hands every member state a seat. The binding rules arrive years later through protocols negotiated under that umbrella. Slow by design, and the slowness is the feature, because it is the only format that has ever got Washington and Beijing signing the same page on anything. The institutions Balakrishnan pointed to are the ITU for telecoms and the IAEA for nuclear. Both do the same thing to the industries they oversee: they convert safety into a certification cost. Certification costs land hardest on whoever cannot yet afford a compliance department. A treaty sold as giving every state a meaningful stake, including those without the deepest pockets, tends in practice to hand the labs with the deepest pockets a regulatory moat. Watch which frontier companies lobby hardest for it. Balakrishnan's three named risks were loss of control over autonomous systems, misuse by rogue actors to build bioweapons, and broad political and economic disruption. His pitch line was blunt: nuclear button, human hand or AI. Turkey's Erdogan made a matching call for a common legal framework from the same podium. The next step that matters is whether either proposal gets drafted into a General Assembly resolution before the frontier labs ship their next generation. Watch: piped.video/WIGwIMSlIvk

At 3:14 a.m. ET on Sept. 27, an address on Bitget's published attacker list swapped 9,999 XRP for bitcoin through THORChain. That came a day after Bitget CEO Gracy Chen formally asked the protocol to refuse service to those addresses, and hours after THORChain said no. Bitget puts the Sept. 24 hack at roughly $387.5 million moved to attacker-controlled wallets. THORChain's reply was to line itself up next to Bitcoin, Ethereum and BNB Chain and ask what any of those networks owe anyone when stolen coins pass through. OKX's Star Xu found the soft spot in that comparison. Bitcoin miners never hold the funds they process. THORChain's validators do. Every asset crossing the protocol sits in vaults controlled by a threshold-signature scheme, which needs a quorum of nodes to sign an outbound transaction. Spread the key across many parties and the intermediary is still there, just with more hands on it. THORChain's own documentation backs him up. Nodes can vote through Mimir to stop swaps on a single chain, stop outbound signing, or halt trading across every connected chain. The switches exist. Nothing in that toolkit filters by address, so honoring Chen's request would mean freezing every user on the affected chain, honest ones included. That same trade-off tripped up node operators after the 2025 Bybit hack, when a vote to halt ETH-to-BTC swaps passed and other voters then reversed it. So the real argument is about where liability sits for a system that custodies funds while they are in transit. Bitcoin can shrug because nobody in the middle ever has custody. THORChain shrugs because its signers decided not to use the controls they already have. Bitget, meanwhile, is paying for the alternative route. It has a live attacker-address API running and is offering 5% of eligible funds that get frozen plus 5% of eligible funds that get recovered, with anything done under court order or law-enforcement request excluded from the payout.

Snowmoon ships under GPL v3, not CC-BY-SA. Adapt it into a movie or an AI-built anime, and the license requires the prompts, scripts and pipeline behind that adaptation to be open-sourced too, not just the code.

The right place for AI training… fitness centre ofc! Only in SG 🇸🇬

Reminds me of dark chill out room where Alex Paterson (orb) & Mix Master Morris dishes out lunar escapade in Heaven UK ‘93 … Good times 😘

Alexandr Wang launched Muse on September 25, pitching it as an AI general manager that turns half-formed wants into executed plans by booking the calls, finding the funding, and tracking the schedule. Agency, not desire, was always the bottleneck.

By 2030 Ethereum verification shifts from full re-download and recompute to PeerDAS sampling plus SNARK checks. Block construction moves from single miner to multi-party builders. Proof of work built a ledger; this setup builds a verifiable computer.

Peter Brandt admits his predicted October bitcoin low missed by months, the real bottom hit in June. He still calls a fresh bull cycle underway, with Bayesian odds favoring BTC above 300k and near 500k by late 2029. His altcoin math: 99% of new tokens end up worthless.

This reminds me of the millennium (Y2K) bug… soon every website will need an overhaul. 😅

… expect more of such incidents 😂

Opus 5.5 lists at $4 per million input tokens, $20 per million output, and $0.20 per million cache reads. Those three numbers hide the part that actually drains a budget, which is that an AI coding task is a loop, and every lap resends the entire conversation so far. Claude Code's own cost breakdown runs the maths. A task that starts at 20K tokens of context and grows to 120K processes about 2.8M input tokens over 40 turns, but only 1.75M over 25 turns. That works out to $1.62 versus $1.02 with 90% of input served from cache, and the same 2.8M tokens would cost $11.20 with no cache at all. No other setting moves the input bill as far as the cache hit rate does. Output is the other trap. One output token costs 100 times a cache read on Opus 5.5, and thinking is billed as output whether or not the summary on screen shows it. So 60K output tokens run $1.20, the price of reading 6M tokens from cache. A model that botches a fix pays the resend again and thinks again, so a retry wipes out whatever a lower effort setting or a smaller model saved. The baseline is about $13 per developer per active day across enterprise deployments, with 90% of users under $30. Give the model a test or a build so it can check its own work, use /clear between unrelated tasks to keep the cache warm on one job, and compare /usage across models on a single real task before trusting anyone's calculator, including this one.

What if the AI boom is just a house of cards? Trillions in debt-backed capex meeting cheap, open models. If inference gets commoditized and massive GPU clusters sit idle, who pays for the depreciation? Are tech giants building a money machine or the ultimate debt trap?

Satya's rule for picking AI stocks: inference is revenue, training is just R&D. Income statement laws of physics always win. R&D eventually caps at 10-20% of revenue. Don't bet on who burns cash training models. Bet on who actually monetizes inference at scale.

China is on a three day Mid Autumn Festival break right now and if history tells us anything it is that this is prime time for protocol exploits and exchange hacks. When the dev team is offline bad actors strike. Makes you wonder how much of this is opportunistic versus an inside job.