0xf35

1/ Some things I've learned recently running coding agents on large-scale projects. Most of this contradicts advice from 6 months ago!

Good morning my loves, happy Saturday. Sorry I've been quiet, obviously been busy, but thought it'd be nice to give you all the details on the multi-strategy absolute return program that experienced the 28% drawdown this year. (1/n)

🦔Microsoft canceled its internal Claude Code licenses this week after token-based billing made the cost untenable, even for a company with effectively infinite cloud resources. Uber's CTO sent an internal memo warning the company burned through its entire 2026 AI budget in just four months. American AI software prices have jumped 20% to 37%, and GitHub (owned by Microsoft) is dropping flat-rate plans for usage-based billing across its products. My Take The AI subsidy era is ending in real time. The same company that put $13 billion into OpenAI and built the Azure infrastructure powering most of Anthropic's compute just looked at the bill from a competitor's coding tool and decided it was not worth paying. That is not a productivity failure on Anthropic's end. Token-based pricing is forcing every enterprise customer to confront the actual cost of running these models at scale, and the number turns out to be far higher than the flat-rate experiments suggested. This ties directly to my Gemini Flash post yesterday. Anthropic, OpenAI, and Google all raised effective prices in the last six months. Enterprises that built workflows assuming AI costs would keep falling are now watching annual budgets evaporate in months. Two outcomes look likely from here. Either enterprises scale back AI usage to fit budgets, which slows the revenue ramp the labs need to justify their valuations ahead of IPOs, or the labs cut prices and absorb the losses, which makes the unit economics worse at exactly the wrong moment. Both paths land in the same place, the numbers stop working, and somebody has to take the writedown. Hedgie🤗

So you want to invest in quantum computing eh? You should probably ALSO invest in quantum resistant tech...like $BB. TLDR: Blackberry doesnt make phones anymore...they do edge AI/physical AI via QNX & secure quantum resistant cryptographically secure comms that are used by @NATO The world will soon realize we need checks AND balances. Graphic credit to @silberschmelzer

@Pivot922 Watch this youtu.be/F3VFYNLFcLQ?si…

By our calculations, $1.33B out of $4.4B of USDe's backing is lending against itself. Here’s the rule we used: Estimated self-lending = gross borrowing against USDe/sUSDe collateral × Ethena’s share of supplied liquidity in that market So if a pool has $556M borrowed against USDe/sUSDe, but @ethena supplies 47.4% of the liquidity, we attribute ~$263M of that as Ethena-funded self-lending. We’re not counting the full amount borrowed against USDe/sUSDe as “self-lending” where @ethena is not the only lender. Using this pro-rata method, we get: - Estimated Ethena self-lending: ~$1.33B - Gross amount borrowed against Ethena assets: ~$1.67B - Difference from pro-rata attribution: ~$336M Sources of data: AAVE: research.yuzu.money/aave-exposures Steakhouse USDtb: #overview" target="_blank" rel="nofollow noopener">app.morpho.org/ethereum/vault… Steakhouse Prime: app.morpho.org/base/vault/0xB… Kamino: kamino.com/earn/lend/ethe… Juplend: jup.ag/lend/ethena/ma… Backing: app.ethena.fi/dashboards/bac…

Trading rules that actually matter: - Never add to a losing position just because “it has to bounce or dump” - Cut losers early. Ignored risk almost always becomes a bigger loss later - Don’t force trades. If the setup isn’t there, do nothing - Preserve capital first. You can’t compound if you blow up - Detach your ego from the trade. The market doesn’t care about your bias Most trading problems aren’t strategy issues. They’re discipline issues.

Favorite / standard AI workflows right now. Notes from the Permanent Underclass. Currently optimized for Codex (preferred agent) 1. Long Running Jobs: TMUX Whips Pointed to Markdown Goal Milestones I more or less gave up on /goal. Maybe they will fix it at some point but it kept saying things were done when they were not done. this also could be "skill issue". So now I come up with a milestone that can be read by a Tmux automator with instructions to work against a burndown list in the milestone and update it every time it does something. Usually set these things in the morning, and before I go home at night I find that automating Tmux is still more reliable than Codex Exec but I'm hoping that changes at some pt bc it is just an insanely jank way to live 2. Favorite Easy API GPU Provider with Bells and Whistles Modal. It's really good! Makes interacting w GPUs far less painful especially for quickly testing things It's also worth $4.5b now so ... fml I think with 90% of these products you'd have been better off just buying their stocks instead of using their products. lesson in there? regardless. 3. OpenRouter ZDR Maxing With A Caveat OpenRouter has a nice provider list that has ZDRs. Which is better than nothing. Obviously sometimes it's worth it to not trust anyone but it's not good for moving fast. Be careful though, some of the OpenRouter ZDR guys don't actually serve the right model so I have a standard smoke test (21 responses to random alpha numeric thing at low temperature) to make sure it's statistically decent. I've seen some crazy crap... quality issues. test your stuff 4. Chatbots are No Longer Instruction Junkies The old way of prompting (do these 8 things and make sure not to skip any steps). No longer works well. This is Known. But many prompts simply never updated. 5.5 chat-latest is a beast. Its latency is crazy. It is good for a lot of workflows. Its web search latency is also crazy. It seems to work best when you prompt it with XML so that is still a thing Also -- an old prompting strategy (ramming PGvector quotes) seems to perform better w new models which are better at ignoring inputs but subtly referencing them without overriding the chat (so if you used to jam vector stuff from authors in, the convo felt very non linear - now they are better at following instructions to reference the material without 5. Deepseek Flash as a Memory Hog Deepseek Flash is good enough to do a lot of really rote tasks for dirt cheap and you can bid it off some decent private providers on OpenRouter / Venice etc. Very good for memory in chat as an async function. (ie categorize what the user was talking about and store the core points) 6. Working out of Wikis: After I realized that LLMs loved HTML format and HTML format is not human readable I started working out of Wikis with lots of formatting and examples I previously mostly lived in markdown out of repos and vscode After each milestone I document everything with HTML tables, mermaid diagrams etc in a nice looking wiki. And I have planning milestones in the planning section of the Wiki that has become my default command center 7. Codex Tailing Codex Every time after a feature is shipped I have a 2nd codex instance tail the docs, find p0s and p1s and enforce linting standards, code quality, latency etc [yes at some point, this just looks like running a company, and I think that's probably the point] . then does worktree code review and then merge each time -- I only do this when I'm really paying attention to something bc both terminals are quasi active [Tank, Cleric set up] 8. Do Only This One Thing Until You Have Achieved Enlightenment This is a ridiculous one. But for things that are mind numbing, like automating financial number extraction I realized that Codex likes to make big batch jobs or long running things that are expensive and usually don't resolve. Then it lies to itself about how good the process was This seems to be basically always true in finance. So now, I just have it go company by company for like 14 hours a day for various tasks (number extraction, transcript analysis, rote stuff) You can leave these things running for 24 hours a day -- I have one job that is 3 weeks in and it's gone from 41% accuracy on extracting all kinds of stuff from company filings to over 85% This is doable now and it didn't used to be. Am sure they will fix this tendency at some point but for now journaling and hegelian evolution w mundane terrible stuff is the jam 10. Deepseek V4Pro Batch Apply Security Audit Codex is expensive/ slow. Can jam every file in DeepseekV4 pro async and say find everything that could be dicey then passing them into Cursor to get another set of eyes on things that is nice and differentiated from Cursor [I started using this again, I figure there is no way that Elon Musk doesn't spray it with free money]. Hopefully the new Grok training is awesome so there's another player wasting money on us 11. Claude Making JSX: Codex, Make This Mock My hatred of Claude as an entity runs deep but he simply is a far better designer than OpenAI (except when it comes to fully copying other ppls UXes at which OpenAI with screen shots is quite good) Do Not Tell claude ot make something beautiful because he has some weird fetish with italics font. "Minimalist" and "match this UX" are good instructions. Codex will "match the UX" with horrible boxes. Claude will get it right. So do the mock in Claude and pass it to Codex 12. Pressure Testing on Private Mode The models gas you up so if you're pressure testing always incognito mode. should go without saying but important 13. Being more subtle with codex skills. Codex Skills are a bit too aggro -- if you overload them with info it has a seemingly nasty effect on context windows and performance. ANd codex uses them to heavily "I am now referencing this Skill" literally every prompt almost. So it's good to put more high level stuff in there that isn't even a skill as opposed to agents.md 14. Clearing Tech Debt 5.5 xhigh is the only model so far seemingly capable of clearing technical debt without making things worse. So important to set aside time for refactoring all the garbage you vibe coded the past 12 months on worse models

Do you get it yet, anon? This stick of RAM weighs its weight in gold. The cost of AI capex has multiplied by 20x due to memory costs. AI build outs are going to begin to fade or downsize because of it over the next year. Memory cost will become its own 🐍ouroboros.

There are very few companies on earth more well positioned than Intel, Sandisk, Lumentum, and Astera.

Longs on @binance and @HyperliquidX quietly overpaid $480 million in funding last year on $BTC and $ETH markets alone Almost every major perp exchange copy-pasted BitMEX’s 2017 emergency setting… and "forgot" to touch it for nine straight years Binance carved out an exception for their own token. Here is why🧵

Jez read Reminiscences of a Stock Operator and shipped it on top Hyperliquid’ tech ( not @Lighter_xyz ) onchain bucket shop. he confirmed it in this thread under @dipsybitsy “yep, fully transparent and the losers who fund it will own.” the bucket shop pattern is 100 years old. customer trades at the listed price, the house takes the other side, nothing reaches the actual exchange. Livermore made his early stake trading against them, then got banned from every shop in Boston for being too good. they died from two pressures: directional regimes that smoked the house, and ticker manipulation where operators pushed the real venue to harvest at the shop. both are live here. neither is solved by stacking five defensive parameters. what is actually onchain: a contract that reads HL’s BBO via precompile and synthesises swaps against a Martingaler LP. zero flow reaches HL’s orderbook. zero funding rebalance. the only feedback loop is parasitic. BBO manipulation attempts to harvest the synthetic spread degrade HL’s book without compensating the makers wearing the jitter. the venue extracts liquidity quality from @HyperliquidX ( fair enough!) the fair launch plus 1000x leverage plus PAPER mint-on-loss design is engineered for initial FOMO. expect a fast first week/month. Jez’ inner circle volume comes in, asymmetric impact taxes the few winners, losers fund the LP and mint PAPER, the chart can look initially great if execute right. this is the bootstrap window the docs explicitly subsidise at 100 PAPER per $1 of LP gain below 2M. Then path dependence will shows up. LP edge per trade is thin: 0.2 bps jitter band plus the convex bite of asymmetric impact on winning closes. order of magnitude, on a $1M ticket the LP edge is in the tens to low hundreds of dollars. one 5% adverse move on $1M of net synthetic exposure is $50,000 of MTM loss. edge accrues like √t under stationary flow. directional regime hits the LP linearly in t. √t vs t means one persistent trend eats months of accrual. the half life of LP equity in a trending regime is short, and it gets shorter as size scales because the OI cap trails realised OI rather than leading it. then the debt queue activates. martingaler issues claims paid by the next loser. once the queue is observable, the EV of opening a new position is negative twice: once because winners get haircut, once because withdrawals sit behind every prior bust. rational flow stops. Queue does not clear. PAPER, a claim on future LP revenue, is then a claim on revenue that must clear the queue first. the cross-venue attack (@NaiveKrypto point in the thread) has its own path dependence. wicking HL spot or perp to harvest the synthetic at the shifted BBO is only EV-positive once papertrade’s OI is large enough that the harvest exceeds the cost of pushing HL. below that threshold, no one bothers. above it, every desk with an HL maker does. self-limiting at small scale, openly hostile at large scale. the steady state is zero, the only question is the path. all bucket shops are zero in short to medium term. the 1900s wave died within two decades, mostly to regulation and to operators eating them. the onchain version skips regulation but keeps the operators and adds a debt queue that hard-codes path dependence into the equity curve. the docs read like a list of every defence shops historically tried, in order. stacking five does not change the shape of the failure mode, it just shifts where the failure surfaces. closing thought. the whole structure is short volatility against directional flow, dressed up as a perpetuals venue. it is the same trade as a Korean retail account systematically selling MRNA noise reversion while their own KOSPI 200 trends straight through the stop. works beautifully until the regime turns, then the half life is a week. clever paper, novel emissions curve, real engineering. the structure is 100 years older than the precompile. Hyperliquid

proud to introduce @papertrade_xyz - a fair-launched, fully-onchain perpetuals exchange built on hyperliquid by @izebel_eth & @blurr -1000x leverage -0 slippage -No funding costs -Self-bootstrapping LP coming soon. learn more at: docs.papertrade.xyz