Hunter Lebow

@united 7/12 SLC->EWR 10:55am incrementally delayed 10+ hours to 9:30pm then ultimately cancelled?? Help make this right!

@Dictatorvoice @ryanpatrickauth @KaracaSerkant Using medium effort level I was disappointed but have been surprised with the quality of product 4.8 Max effort can ship. Visualizations definitely got a big upgrade.

Anyone have a framework or best practices for selecting effort level when using Opus 4.8? And if token management is removed from the equation how would that change the methodology?

First day using #Opus 4.8…the visualizations are powerful, but the text output I’m getting has been materially worse than 4.7. Anyone else disappointed?

@Jmoon_174 Agreed. AI buildout accommodating worst-case token efficiency, which is reasonable. The dust will settle eventually though and an equilibrium can be reached

@hunterlebow part of it is that the tooling to actually measure per-call waste in production barely exists yet. you kind of have to build it

Building with AI is as inefficient as it will be. Disciplined token management is hardly a consideration in practice right now across AI enabled companies. Once efficient systems to build with AI are standard, compute demand will stabilize. We’re in the (expensive) sandbox phase

@mattkalish 🧎‍➡️

@hunterlebow I have no clue I just make internet content now

@mattkalish How do you think about Kalshi’s Liquidity Incentive Program. Will DraftKings encourage this on DKPredicts?

@orrdavid Agree. Building with AI is as inefficient as it will be. There is little token management in practice right now across AI enabled companies, it’s all an expensive playground. Once efficient systems to build with AI are ubiquitous, demand will stabilize. Could take years though.

My AI bear case, really starting at point 4: 1. Lots of people and companies were still skeptical on AI in Q4 2025. Sentiment began changing a lot in Q1, 2026. They began to realize that AI really could do wonderful things for their business. Ken said this in some interview. I know several people personally who are saying this. One fund manager, who is strong at fundamental analysis, told me the same thing - a year ago he thought AI was stupid, today his jaw drops on the analysis it does. And company analysis is much more abstract and less rules based than most white collar jobs, like accounting and law. And AI is much stronger for rules based work. We just hit a new threshold of performance that really does threaten humans in almost all white collar jobs. 2. This new "high enough level of performance" suddenly increased AI demand. Compute rental prices skyrocketed. This in turn causes the hardware prices to skyrocket. Hyperscalers won big, so did hardware companies. 3. Companies are justified paying these high prices. The immediate economics are probably good, but even if not they can rightfully assume they will be looking out a few years. Wait, but isn't all of that a bullish thesis? Yes, but short term only! What happens from here? 4. The mass white collar replacement will only happen once. It's a single shock. 5. People are still terrible at using AI today. It's like the internet in the early 90s, a clusterfuck. Companies will get *WAY* better at actually using AI. They will increase efficiency, reduce lots of wasted compute. Right now they are building several competing, inefficient systems. Eventually they'll wind up with one wonderful system. Once set up, demand for compute drops. 6. Compute power is getting way stronger every year. Point being, in another couple of years companies will only even need the same amount of compute as they do today. Or maybe a bit more, but not that much. But the amount of compute available will double because the hardware is getting so much more powerful. And then in another couple of years it will double again. This is wonderful for the companies using AI. It is terrible for the companies making AI. It seems like a downright terrible business. 7. While some AI applications will use way more compute in the future, and long term I am optimistic on what AI does, the big one time wave is what's happening above. The one time wave is like 90% of the work AI will do over the next 5-10 years. Think of all the boring repetetive tasks that are going to be automated away. Even if AI creates a lot of new work over the same timeframe, which probably, it will be tiny compared to the knowledge work being destroyed. It seems a quite possible boom/bust cycle. The hardware is too commodity like and too much of it will be ordered. That's true even if AI will do amazing things looking out 20 years.

Reimagining the AI Trade $RDDT, $GOOGL, $TSLA, $META are the bronze intelligence layer AI Labs are intelligence manufacturers and token wholesalers $SNOW, Databricks are the token retailers $PLTR is the doordasher of tokens

@orrdavid Interesting seeing VIX / VIX3M fall and RSP / SPY rise with this sharp move. The move is being treated like noise, not a liquidity unwind

Big yen move.

@kevinxu Survivorship bias and narrative fallacy

i traded $35k to $10m and never read a single investment book. theres a lesson here.

The market is overestimating how AI negatively impacts softwares moat. I’ve yet to see a single mainstream competitor emerge from vibe coding and AI tools. Also, quite ridiculous companies like $UBER are bucketed in this category. Buy.

@Mr_Derivatives $SLV / $GLD

Entering 2026, if you have to short one name, just one... What would it be? This is a safe space bears.

As much as I am a long-term $HOOD bull with a multi 6-figure position built in Summer 2024, there has been objectively bearish market structure since Oct 6th (lower highs, lower lows), and no one is admitting it. Watching for an h-shaped bleed to 200MA until trend is invalidated

$HOOD S&P500 inclusion gap-fill to $104.98 is inevitable. Looking to add more once that has been achieved and resume higher.

@vladtenev is the Gen-Z's Steve Jobs

#ShedeurSanders was a projected top pick in #NFLDraft2025, but ultimately slid to the 5th round, the 144th overall pick. Here’s what it taught me about B2B SaaS 🧵…

@TobyPhln @elonmusk Arrays are stored in contiguous memory while linked list node objects are not stored contiguously in memory nor gaurenteed to have any relative proximity in memory, making linked lists often less cache friendly than arrays

Sounds enticing? Join us! Looking specifically for experienced C++/Rust backend engineers for grok.com and the API in the bay area: job-boards.greenhouse.io/xai/jobs/47005… The first interview only takes 15 minutes. Here are some of my retired questions for this role & interview: - Assume you want to write your own string struct. What fields would you define on the struct? - What are move semantics and why is moving a string faster than making a copy? - Explain the differences between preemptive and cooperative concurrency. - Where is the thread scheduler defined and when does it run? - Why is iterating over an array faster than iterating over a linked list even though both operations are O(n)?

“Catch a man a fish and you can sell it to him. Teach a man to fish and ruin a wonderful business opportunity.” -Karl Marx This is the best argument against vibe coding. Learning how to fish has never been easier, being caught fish has never been easier, too. Use AI wisely.