Maxim @mximaf

In the last couple of decades, taste and bias were something you kept in the background, or even had to overcome as a subconscious urge or gut feeling. The instruction was always to be objective, follow a framework, and implement best practice (especially during my time in consulting). There always seemed to be a right way of doing things. That trend is shifting. Over recent decades, and especially with AI in the last few years, building and executing ideas has become much simpler. Lower barriers let people who did things differently break through and succeed, sometimes against conventional dogma. A few reasons for the shift: (i) Listening to your gut, taste, and bias lets creators lean into their strengths. There is comfort in not constantly having to override that instinct. (ii) On the customer side, authenticity and originality are highly valued. (iii) Frameworks, best practices, and expert opinion are becoming less valued. Predicting the future is harder than we thought. As information and expertise become widely accessible, the “right framework” differentiates less, and judgment and taste matter more. I’m not arguing against objectivity. But we’ve spent too much time trying to eliminate taste and bias rather than learning when to trust them.

Early capital should address a business’s greatest “white-hot” risks threatening growth in the cheapest way possible. There’s a wrong way of “de-risking” that looks like expensive theatre - throwing money at unknowns. Hiring excessively. Building 20 solutions for non-prominent problems. Raising bigger rounds. These convert one risk into a higher bar one must clear. Real de-risking is an experiment whose result provides information that either increases conviction or requires pivoting. If no outcome kills the status quo, it wasn’t a true test. A good historical example is @KleinerPerkins’ approach for Genentech. Genentech’s white-hot risk was scientific. Instead of building a lab and hiring 10+ scientists, Genentech contracted its core science to university labs when it was unanswered whether microorganisms could produce a human protein. Once answered, it stops being a question and becomes an asset (in Genentech’s case, IP). Then you own it, and that’s where the compounding lives.

Musk's version: learn everything about the Soviet engines, then subtract your way to something better, yet simpler

While all consultancies do that - copying best practices caps your ceiling on solution quality. At large, you can copy a 7-8/10 solution you don't own or build from the ground up that might land you a ~6/10 solution. Over time, however, solution's ownership wins over pure quality, simply because full ownership is what makes iteration possible. You can't iterate on a design when you don't understand the separate drivers (the micro-solutions) that produced it. You can only replace it wholesale. We usually want "now" and "the best", which lures us into copying from benchmarks . Iterating over time wins (on quality and customizability). inspired by @tobi's take from @davidsenra's interview

True. Socialism concentrates power around politicians through incremental laws and regulations. Capitalism shifts power toward wealth creators. However, only capitalism is truly egalitarian and can gives new wealth creators a seat at the table.

Hayek-Club Weimar @WeimarClub

a month ago

Sozialismus ist das perfekte Geschäftsmodell für Politiker: Er erzeugt Abhängigkeit, und Abhängigkeit erzeugt Wähler. Wer vom Staat lebt – durch Transferleistungen, Subventionen, „soziale“ Absicherung, planwirtschaftliche Privilegien – wählt denjenigen, der die Hähne aufdreht.

185 2K 7K 715K 3K

I think founders should steal from the PE diligence playbook. Day 0 of a diligence: call customers, competitors, former management of the asset you’re diligencing. All this in purpose of, getting smart on the market and getting VERY deep on the problem you’re addressing. Founders should do the same. Understand the problem before obsessing over the solution. Other playbook elements founders could steal are - (i) mapping key drivers and KPCs (what customers value, in what sequence), (ii) creating interview guides and surveys that actually get you the data that deliver an insight on the market (e.g., product market fit, target customer (or even ideal customer profile)) Where it breaks: no direct comps for a novel problem. So competitive analysis becomes problem comparison.

Travel recs are a trust problem, not an information problem. Hot take, but travel content is mostly worthless because it’s optimized for everyone. And a recommendation that works for everyone doesn’t create a strong pull on a personal level. (Plus, “Top 10 things to do” solves information - it’s already solved) “Trust me, go here. You’ll love it” - that is somewhere I’d go.

I'm increasingly seeing why planning (especially long-term planning) is really just accumulation of assumptions. That's why I love this from @vkhosla: "I always tell our CEOs: don't plan. Keep testing the assumptions and iterating." The future isn't predicted (historical patterns only hint at potential outcomes). It's discovered by iterating through the assumptions

I increasingly think careers can be plotted on two axes: breadth/depth and agency. And they’re not independent - the deeper you go into one problem, the more agency you tend to have. Founder - maximum depth, maximum agency Early startup - high depth, high agency PE/growth - medium depth, medium-high agency VC - enormous breadth, moderate agency Consulting - enormous breadth, relatively low agency

Checked out @gonka_ai, a decentralized AI compute network (GPU owners earn tokens, devs get cheap open-weight inference through an OpenAI-compatible API). In simpler terms, anyone can have their agent run on Gonka compute to access open-source models cheaper. On the surface I really like the idea, but this brought up two questions / bottlenecks for decentralized open-source inference. 1. Open source vs. frontier: A more macro driver here. Open-weight models have narrowed the gap across the board, and especially on coding benchmarks and remain x times cheaper to run. Being 'good enough + cheap’ seems to be a solid model here for compute market share. The open question, though, is whether today’s “good enough” holds into the future. Open source has traction precisely because the quality delta is realtively small. If it widens, market x use cases start tipping back to closed models and that's plausible. Plus, 100-step agentic runs compound failure. 2. Centralized vs. decentralized inference. I’m less certain about this one. Decentralization (Gonka is built on a native crypto chain that helps incentivize compute providers) promises censorship resistance and less risk from ‘gatekeepers’. The team behind Gonka definitely go hard on pulling the “ideology as means for marketing” lever (e.g., freedom to the users and open sourcers, etc.). The issue is that centralized aggregators (@OpenRouter, @togethercompute) match or beat decentralized options on price, speed, and user experience (cloud infra + it’s just easier to live without crypto-related overhead + everyone uses the same API format, so switching providers is super easy). This creates market decentralization, even if the infrastructure isn’t decentralized, and it keeps any single provider in check. Also worth saying out loud: decentralized doesn’t mean private. You’re trusting more parties with your prompts, not fewer, just with diffuse accountability. So the real question for these networks isn’t about being closed or open. It’s about what value add they offer (if it’s not price, latency or user experience). (My guess: there will be some sovereignty demand from anyone who doesn’t want to depend on US-controlled infra; same pull that’s driving open-weight model development) Let’s see how this plays out

Open source becomes both the floor and the ceiling for AI intelligence. Closed labs won’t win on raw intelligence (or marginally). They’ll win on the ecosystem: harnesses, and tools optimized for specific use cases (design, coding, app dev, SMM, and so on). So it becomes a game between two builds: Open source intelligence + an ecosystem of incremental apps vs. All-in-one (Claude/Anthropic, OpenAI) (Apple-esque), and potentially hardware stacked on top of either side

Dean W. Ball @deanwball

2 months ago

Some observations on Kimi: 1. It's a very good model! I don't think its performance can be explained away by distillation or anything like that. In agentic coding sessions, it seems pretty much on par with the best public models of Q1 2026. In my fairly limited use, it also

2K 936 8K 11.3M 9K

“The toys were fine. But we always wanted to play with the boxes” resonates quite well. To create, to rebuild is truly human, and making it easier is near-sacred. open.substack.com/pub/a16z/p/the…

AI and robotics offer massive technical headroom for optimization on every level (progress acceleration), but change management (ie how nations, societies adapt & coordinate) remains the dominant source of near-term progress downside risk

bayes @bayeslord

3 months ago

x.com/i/article/2072…

79 248 3K 1.6M 9K

@kadmitriev Бензин верни сначала у себя в песочнице прежде чем на других хаять

The average life expectancy of a new Russian recruit—from arrival at a training ground to death in a combat zone—lies somewhere between 10 days and three weeks. Once sent onto the battlefield, they survive an average of 20 to 35 minutes. @peterfrankopan foreignpolicy.com/2026/06/25/rus…

Last night in Kyiv might seem “typical” in post ‘22 reality, however, one of the holiest sites of Orthodox Christianity was hit - something dear to the two populations and even to country leaders themselves. Fascinating how sacred beliefs, fundamental to so many, could be bent

@defivas No London on the list?

@garrytan @grok what is 10k LOC a day?

@alexeyguzey love the Moscow subway selfie profile pic

@chamath I like how @chamath looking at a chart with applicants and says that’s the number of people granted visas. The number of visas granted doesn’t change - it’s capped at 85k