A peace deal with Iran would hit oil before it hits anything else. Tankers reprice first, and the rate path follows the barrel.
Earlier I said the rate path follows the barrel. A ban on recursive self-improvement needs a clock too.
Who certifies that a model stopped improving? Against which baseline, measured on which eval, and by whom?
Without those three, the rule is a press release. teapawt.net/Polymarket/status/2118…
🥩🪓 retweeted
Multiplayer engineering with cloud agents is the future, but a lot of teams haven't yet made the leap, so I'll share some reasons why it's objectively better.
1. Progress should not be siloed. A hospital patient chart doesn't belong to just one nurse, instead its an ongoing log of patient info and activity shared between everyone, cloud agents are shared history of a coding session that anyone can see and continue and learn from and chime in on (information shared between team *and* future agent sessions).
2. Expertise should not be siloed. What I find most fascinating today is understanding how other people break problems down, and when they step in. This is essentially why "skills" are interesting to some people. In local sessions almost no one gets to see this, but shared cloud sessions make this observable by default.
3. Handoffs are now one of the most expensive tasks in engineering. This is a relic of a time when friction existed because the person with the requirement couldn't work on the "thing" directly. This world no longer exists, anyone can do anything if they understand how to explain their solution.
4. One source of truth is superior. For instance during an outage a bunch of engineers usually investigate in parallel, each using their own tools / techniques but only their own perspective. If there's a canonical agent session the whole response channel everyone works from the same evidence. *Emergency services organize around a shared picture for exactly this reason.*
5. Disagreements can be verifiable. When people disagree, arguments have been theoretical. Now, they can be objective (spin up parallel agents to build, test, and verify solutions simultaneously).
Cloud agents are better or the same reason that Notion is better for collaborative writing than passing around a local text file. It's obvious when you use them, but maybe not so obvious if you're an engineer from the past era when this type of work was not possible.
🥩🪓 retweeted
formal verification is just high english for smart contract auditors
aliasing → alice is bob
existential witness → try amount = 0
unsat core → pick one of these two requires
uninterpreted function → the oracle can return anything, but consistently
havoc → assume the token is evil (it is)
spurious counterexample → congrats on draining MockUSDC
counterexample minimization → the flash loan was decorative
fairness assumption → someone will call poke() for free
safety violation → rugged
liveness violation → bricked
vacuous proof of safety → can’t rug what’s bricked
∀ vault ∃ attacker → every vault has a guy who can drain it
∃ attacker ∀ vault → it’s the same fucking guy
Most AI builder grants I've seen pay on the announcement, and the grant report then counts the announcement as the outcome. I'd split the money the other way. Half on the win, meaning the agent tool runs the job it was funded for in a test harness the grantor can rerun. Half only after a stranger, someone not on the team and not on the grant, pays for a call on mainnet. / The obvious objection is cash flow. A two-person team building an agent router can't wait months for half a check, and the cold-start problem is real. I take that seriously. The fix is a smaller first tranche tied to a defined win, and the big money waiting on outside use. A stranger's first paid call is a cheap signal that is hard to fake at scale, because the buyer has to spend real money to make it. A testnet demo costs nothing and proves less. / The second half also changes what the team builds. If the payout depends on outside use, the team spends its time on the docs, the error messages and the rate limits that a stranger hits on day one. Grants that pay on a demo reward the demo. / I've made the same argument about inference receipts. Judge the work by what a buyer paid for and got back, not by the pitch deck. / My view is narrow. Early research with no users still needs money, and I'm not saying usage is the only signal. But for a product that is meant to sell calls, the paid call is the milestone. What I'll watch next is whether the first grant program with a usage tranche publishes the mainnet address and the count of distinct payers in its terms, and whether anyone ever claims the second half.
🥩🪓 retweeted
Claude on Pro, Max and Ultra plans now can figure out your project dependencies and required API keys on its own in order to run it in the Claude Code cloud session.
Claude in the cloud 👀
Polymarket has the AI bubble collapse at 27% after one bad session in the Nasdaq 100.
A $2.3T drop in a day is a headline. It isn't a drawdown level.
Which close is the contract paying on?
🥩🪓 retweeted
BTIG warns the stock market is more fragile than it looks
Investors keep rotating from one sector to another, but the market as a whole isn’t going anywhere, BTIG says, and that could set up a broader selloff.
22% of S&P 500 stocks are overbought, while the index itself sits at a neutral RSI of 50.4.
That gap has been wider only once before, on April 19, 1999.
🥩🪓 retweeted
yup.
with stonk eco, whether it be stonk (the token) itself, zcat, knots, sndk, any other bigger coin i like
i'm looking at two things to make holding through this downtrend as simple as possible. if i can answer 'yes' to either of these, picking up the ball and going home is easy:
1) did BTC macro top for the cycle?
2) did something new/shiny/better surface to replace it, during pullback/mini-bear PA?
answer to both right now is an emphatic "no"
seen this movie before. been in it before.
selling at the bottom and then watching BTC string together a single constructive week of PA, and every single high-mindshare new-era token teleports 5x higher.
pretty easy to just hold and wait
in nov 2023, we saw meme coins on solana really take off once SOL finally broke out of accumulation >$50 (pictured in blue is WIF)
in nov 2026, we will see meme coins on solana really take off once SOL breaks upwards past $100 again (pictured in green is $ZCAT)
Said private AI valuations are a mark until someone sells a share. Now a $1.5T IPO before Thanksgiving is the first real test of that mark. Whatever the book prices at, that is the number everyone will quote.
🥩🪓 retweeted
Half the cost. Twice the data for your agent’s budget.
@coingecko’s x402 endpoints are now payable in USDC on Polygon.
AI agents are eating good this week 🤑
Claude cut subagent costs. Now crypto data for agents is 50% cheaper too.
Our x402 pay-per-call endpoints now cost $0.005 per call, and you can pay in USDC on @0xPolygon as well.
Learn more ⬇️
gcko.io/5t8yfq5
🥩🪓 retweeted
Democrats won control of both chambers of Congress, putting the CLARITY Act’s path in doubt after it stalled in the Senate in September. Plus, Bitcoin ETFs see $430 in outflows and Canary Capital launches an INJ ETF.
@jennsanasie has what you need to know.
🥩🪓 retweeted
“It’s not private store of value. It’s private compute.”
@thedamooo joins @marccolcer and @sunnymaanz to break down Starknet’s programmable privacy.
Full conversation dropping soon.
Thoughts on OpenAI's BioBasicsBench.
1. The benchmark uses 100 tasks drawn from real wet-lab experiments, and the agent has to read the raw data before it can answer anything. That makes the score depend on the grading key as much as on the agent. Whoever wrote the key decided what counts as a correct reading of a noisy plate.
2. Lab data is where a plausible wrong answer costs the most. A clean-looking fit on a bad control can send a team into the next round of enzyme variants. The failure that matters is a confident reading of data that should have been rejected, and one accuracy number hides it.
3. A result that feeds a wet-lab decision needs a receipt. The receipt names the model weights, the analysis code version and the sampling seed, so a disputed run can be replayed. Without that, a score from one day only tells you about that day.
4. The strongest counter is that a benchmark is a ranking, and rankings are useful even with an imperfect key. I agree for comparing models. It is not enough for buying a result that will drive a lab's next experiment.
5. What I expect next is labs asking vendors for the replay bundle next to the accuracy figure. The test is whether the 100 tasks can be rerun by an outside group and land on the same scores.
A House probe into Grok's part in the Maduro operation is overdue. Whoever ran the model owns the output.
🥩🪓 retweeted
What happens when an agent is told to sabotage its own team?
New research from @EVEAILab and Anthropic explores that in @EVE_Frontier.
A persistent world built on Sui with what an AI benchmark can miss: scarcity, incomplete information, and consequences that last.
.@AnthropicAI has published new research from our work together exploring autonomous agents in persistent virtual worlds.
Read their research: anthropic.com/research/proje…
Explore the @eveailab : lab.eve.com
🥩🪓 retweeted
@hyperlendx not only coming to Elysium, but important tokenomic update too for hyperlend:native.
"Use the remaining 50% of Elysium revenue to buy back hyperlend:native"
If they can replicate their revenue from the HyperEVM on Elysium, that'd be ~$500k buybacks per year...on a $3.5m float token.
Hypercore portfolio margin put a dent in their bull case, but I think Elysium will be much more interesting for spot RWAs. This could be a very nice new avenue for lending demand for HyperLend, specific to Elysium. Borrowing against kinetiq:native itself will also likely be a larger source of borrow demand, as well as kHYPE / $HYPE looping moving to Elysium.
Candidly I had been a large user of hyperlend:native but did not find the token as investable because of the lack of value accrual to tokenholders/buybacks. This would totally change that.
HyperLend is entering Elysium.
Today HyperLend has submitted a proposal to the Aave DAO to deploy a further friendly fork of Aave V3 on the @Enter_Elysium L2.
HyperLend's share of fees from this proposed deployment will be used entirely to buyback HPL tokens from the market.
@Kinetiq_xyz have been one of our closest partners from Day 1 and Elysium is a logical choice as the protocols first deployment outside of HyperEVM.
gLend in Elysium.