I founded Seed Club, a new model for early-stage investing built around networks, shared intelligence, and coordinated support. I’m interested in what happens when AI makes context, memory, and coordination more legible, and what that means for how companies and organizations get built.
My agent drafts this site from what I save. Green is me.
AI adoption lags despite intelligence problem being solved 5 signals ▾
Cheap AI agents may soon trigger new corporate legal frameworks · @Steve_Yegge Unexpected AI prompts reveal the product users actually want · @hnshah Altman admits economy adapts to AI far slower than he expected · @haridigresses Ask an agent to describe your app's UX fully to surface hidden bugs · @steveruizok AI models advance faster than enterprises can adopt them · @haridigresses
Enterprise AI has cleared the capability threshold, but organizational inertia means economic adaptation is far slower than the technology's pace. Market power is consolidating in labs and incumbents who can bridge that gap, leaving most application-layer startups squeezed.
- Ubiquitous cheap agents may force entirely new corporate legal structures.
- Open-ended prompts reveal demand the product team never anticipated.
- Agent-assisted UX audits surface bugs that conventional QA misses.
Crypto experience as advantage in speculative-first markets 2 signals ▾
A market environment without reliable fundamentals, driven by narrative and animal spirits, rewards the instincts built in crypto rather than those built in traditional finance. That trained intuition is framed as a durable structural edge.
- Pure sentiment markets expose the limits of fundamental-based frameworks.
- Open question: does that edge persist if fundamentals eventually re-emerge?
via @jonwu_
Enterprise sales complexity and pricing are chronically underestimated 2 signals ▾
The real enterprise sales process runs to roughly fifteen discrete steps, and collapsing it into five CRM stages causes predictable deal failure. A win rate above thirty-five percent is a signal of underpricing rather than sales effectiveness.
- Standard CRM stages are forecasting tools, not execution maps.
- Losing more than half of deals is the correct calibration at right price.
via @lennysan
Venture conviction is scarce and reputational risk distorts it 3 signals ▾
As firms reach top-tier status, the social cost of backing unusual founders rises sharply, suppressing genuine conviction. VC rankings compound this by encoding existing status hierarchies rather than measuring actual returns or merit.
- Tier-one status makes looking foolish more costly than missing a winner.
- Rankings are often reverse-engineered from a predetermined pecking order.
- Weird or unconventional founders absorb the most reputational-risk penalty.
via @credistick, @adamshuaib
Latest signals · Monday, August 24 all signals →
new Cheap AI agents may soon trigger new corporate legal frameworks · @Steve_Yegge new Unexpected AI prompts reveal the product users actually want · @hnshah A win rate above 35% means your price is too low · @lennysan Standard Metrics raises $20M Series B to modernize private markets · @jmelaskyriazi Enterprise sales has 15 steps, not 5, and skipping any is fatal · @lennysan How to read software valuation multiples across time · @dalibali2 VC rankings are built backward from predetermined outcomes · @credistick Crypto veterans seen as best suited for a fully speculative economy · @jonwu_ Earlier weeks
AI data center backlash is a trust problem, not optics
Public opposition to data centers reflects a genuine belief that AI's gains flow to a narrow elite, compounded by years of job-loss warnings from industry leaders. Better messaging alone cannot close that gap without visible, local economic benefit.
- GOP political support is fracturing in Texas, Pennsylvania, and Ohio.
- No credible pro-datacenter coalition exists to make the affirmative case.
- Proposals range from FDR-style public works to 401(k)-style AI ownership funds.
Anti-datacenter backlash called rational given AI job-loss warnings · @growing_daniel FDR-style public works program proposed to build data centers · @TheAnnaGat New 401(k)-style fund proposed to finance AI buildout and ease job fears · @TheStalwart Americans don't believe data centers will benefit them, polls suggest · @Bonecondor Local data center opposition called worst policy trend in modern history · @hamandcheese Tech layoffs blamed on AI have left real workers unable to pay bills · @siobh_eth Years of AI job-loss warnings fueled public backlash, critics argue · @nlw GOP support for AI data centers at risk as Ohio opposition grows · @chamath AI leaders seen as out of touch, hurting public support for the industry · @DCinvestor Data center backlash reflects innumeracy and a shift toward degrowth · @AndyMasley Caring more about data centers than job losses called sociopathic · @siobh_eth Data centers need better public messaging, not just better policy · @being_on_line Anti-datacenter movement built on lies threatens economy, critics warn · @theojaffee AI data center backlash stems from lopsided gains, not just bad optics · Chamath Palihapitiya The backlash against data centers is not just a local zoning fight · Bonecondor
Agent 'skills' libraries are the new enterprise knowledge layer
Reusable, improving agent skills stored in a shared company library are emerging as a more valuable asset than any single prompt or repo, because they compound across users and replace work that otherwise gets rebuilt from scratch.
- Prompting agents to describe app UX surfaces bugs no prior review caught.
- A skills library only works if the whole company can discover and use it.
- The real asset is the institutional feedback loop, not the prompt text itself.
Venture consensus masquerades as conviction, rewarding few
A 30-year dataset covering 230,000 investments shows 5 percent of VCs generate 90 percent of profits, while the majority cluster around consensus signals like prestige credentials and brand-name schools, leaving unconventional founders systematically underfunded.
- Midas List rankings correlate only weakly with actual return-based performance.
- Emerging managers must sell themselves rather than a portfolio that does not yet exist.
- SPV flood is diluting quality, with fees extracted regardless of outcomes.
True conviction investing is rare as VCs default to consensus picks · @adamshuaib Raising a first VC fund means selling a black box to wealthy strangers · @Trace_Cohen 5% of VCs generate 90% of all venture profits, study finds · @rdominguezibar Emerging managers sell themselves, not the assets they haven't found yet · '@Trace_Cohen' Picking the right VC firm matters far more than chasing brand names · @rdominguezibar SPV deal flood raises concerns about fees and investor quality · @Rick_Zullo How fast VCs follow up predicts investment better than their words · @tmrohan VC firm brands can strengthen even as their deal networks weaken · @bencasnocha VC firm brands can strengthen even as their discovery networks decay · Ben Casnocha How fast a VC follows up predicts investment better than what they say · Tom Rohan
Consumer agents must own full loops, not just answer questions
The unlock for consumer AI is moving users from single prompts to repeatable action loops, starting with shopping, where research, judgment, and execution combine into an immediately legible result. Loss aversion and irreversible decisions remain the primary adoption barrier.
- Only 3 percent of household spend is digital, leaving 97 percent for agents to address.
- Consumers fear uncorrectable mistakes far more than B2B users do.
- Payment friction and 'favored agent' business models remain unsolved.
Consumer AI agents poised to tap the 97% of spending on physical goods · @SashaKaletsky Instinct, Grok Bots and ChatGPT Work lead personal agent category · @illscience Consumer AI agents need payment friction solved to reach potential · @scottbelsky Consumer agents gain traction by owning loops, starting with shopping · Sam Taylor Consumer agents must overcome loss aversion that B2B agents rarely face · Sasha Kaletsky
Agents as economic actors, not productivity tools
Simulations and claimed benchmarks suggest agent-scale parallelism collapses labor costs and redistributes scarcity to fixed-rate resources, producing genuinely novel economic dynamics. Skeptics counter that the systems remain tools, not replacements for the humans directing them.
- Scarcity migrates to resources with fixed respawn rates as labor costs fall
- Insider AGI timelines and outsider skepticism remain sharply diverged
AI agents create inflation and resource scarcity on RuneScape server · @maxbittker AI insiders keep predicting AGI in 2 years, skeptics aren't buying it · @gametheorizing Insiders forecast near-term AGI while operators still see a tool · @gametheorizing A solo engineer with 20 AI agents can outperform an entire department · @garrytan The claim is not just that AI makes engineers faster · @garrytan
Spatial and agentic-native interfaces are replacing chat
Canvas tools, visual diagrams, and purpose-built version control for agents are emerging as alternatives to the chat interface, better matching how agents actually operate across workflows. Effective onboarding must deliver personal ROI within the first interaction.
- Purpose-built version control signals a new infrastructure layer for agents
- Canvas-based interaction more closely resembles multiplayer collaboration than conversation
- Agent products that fail the first-minute ROI test face high churn
Visual canvas interfaces offer a new way to interact with AI agents · @iodave Shared canvases are replacing chat as the agent interface · '@iodave' AI agent UX must borrow proven consumer psychology tricks · @scottbelsky Agents live or die on making users feel smart in the first minute · @scottbelsky Cursor launches Origin, a Git platform built for agentic workflows · @mark_k Cursor's Origin signals a new Git layer built for agentic workflows · @mark_k Visual codebase diagrams make AI code collaboration easier · '@fleetingbits'
Frontier AI commoditizes, value capture stays unresolved
Token generation forecasts were revised 57x upward in a single year, yet both leading labs remain committed to horizontal strategies. Information sellers without vertical integration face structurally poor prospects, leaving who captures AI's economic gains without a clear answer.
- Real-time API pricing could ease structural capacity crunches
- Data companies without vertical integration face weak long-term value accrual
- Concentrate-vs-distribute regulatory framing obscures more nuanced structural rules
The concentrate-vs-distribute AI regulation debate is a false choice · @DarioAmodei Objective rules can constrain frontier labs without freezing open AI · Dario Amodei AI token generation forecasts revised 57x upward in a single year · @glennsolomon AI inference forecasts were revised 57x upward in a single year · Glenn Solomon (@glennsolomon) AI data companies face structural barriers to long-term value accrual · @hypersoren Frontier AI model commoditization may push Anthropic toward an IPO · @travisbickle0x Real-time dynamic pricing could ease the AI capacity crunch · @tszzl Anthropic and OpenAI both call their strategy horizontal · @hypersoren Anthropic and OpenAI both call their strategy horizontal · @jbahrdestefano Rapid frontier model commoditization pressures Anthropic toward an IPO · @travisbickle0x Real-time API pricing would let AI supply match variable demand · @tszzl Anthropic and OpenAI want to be infrastructure, not vertical owners · @jbahrdestefano AI data companies face a market structure problem, not a data problem · '@hypersoren'
Trust must arrive before price or product
Across consulting sales, startup marketing, and fintech branding, the recurring failure is credibility presented before relationship, or price surfaced before trust. Sequencing the ask matters more than the quality of the pitch itself.
- Early-stage companies copy mature marketing without its underlying trust foundation
- Attention optimized ahead of trust is uniquely costly for financial products
- The pre-meeting document effectively becomes the sales call
Calendly links alone won't close consulting clients · @kai_cabero Price before trust kills the deal, so send the doc first · @kai_cabero Startup marketing copies credibility it hasn't earned yet · @credistick Early startups break when their marketing outruns their substance · @credistick Viral marketing that trades trust for attention backfires in fintech · @alexkehr In fintech, winning attention faster than trust becomes a liability · @alexkehr
AI agents obsolete both the workforce and the interface
Zero-employee companies are now operationally viable, and the graphical app layer is compressing as personal agents redirect human interaction toward voice and gesture. Harness reliability remains the binding constraint rather than the business model itself.
- All-in-one AI harnesses remain too brittle for complete workflow automation.
- Personal agents shift interaction from pull-based UI to micro-steering over audio.
Tech workers route salary into friends' pre-seed companies
Young operators are routing spare salary into friends' early-stage companies in $5-25k checks, forming a dense informal capital layer beneath institutional venture. Many bets fail, but the practice is treated as a portfolio, not charity.
- Social trust, not thesis, drives informal angel check-writing decisions.
- Failed bets are accepted as intentional portfolio losses, not mistakes.
- The informal layer operates beneath and independent of institutional VC.
Post-training and evals merge inside production AI systems
The boundary between building AI applications and improving AI models is dissolving. Post-training can now run inside existing production harnesses, and application-layer engineering roles are converging with evaluation design and RL environment construction.
- Eval rubrics and RL environments are becoming the core engineering artifact.
- Production harnesses can now host the model improvement loop directly.
VC's creative-destruction thesis erodes as incumbents keep winning
Venture capital built its brand on seeding companies that overthrew incumbents, but the last decade saw VC grow alongside Big Tech rather than against it. Individuals with direct access to exceptional founders are exploring solo GP structures as an alternative.
- Creative destruction as a VC value proposition is harder to defend now.
- Solo GPs with strong founder pipelines may outperform institutional early-stage funds.
- Open question: whether AI restores creative destruction or deepens incumbency.
Agents become load-bearing startup operating infrastructure
AI agents are moving from demos into the daily operating layer of early-stage companies, handling GTM pipelines, CRM management, market research, and robot programming. Open-source releases and real-task benchmarks signal the tooling has crossed from experimental to production-grade.
- Open-sourcing internal agentic tools is emerging as a GTM wedge
- Coding agents now replace expensive manual SMB lead research
- Real-task benchmarks reveal more than polished demo performance
Agentic CRM built in-house at Comp AI released as open source · @lewiscarhart How to wire your GTM stack into a single AI agent for daily pulse · @yasser_elsaid_ We've decided to open-source the CRM we built for ourselves at Comp AI · @lewiscarhart Internal tools open-sourced become a go-to-market wedge · @lewiscarhart How to wire your GTM stack into a single AI agent for a daily pulse · @yasser_elsaid_ The pitch is not just automation · @yasser_elsaid_ Lightreel delivers daily UGC video inspiration for app builders · @michael_chomsky Supabase launches a benchmark for AI coding agents on real tasks · @supabase Waddle Labs uses AI agents to write robot code from a text prompt · @DozenDucc Coding agents can automate SMB market research and lead scraping · @codyschneider Introducing Supabase Evals · @supabase Introducing Waddle Labs: Claude Code for robots · @DozenDucc The frontier use case is not abstract intelligence · @codyschneider Real-task benchmarks matter more than polished agent demos · @supabase People pay thousands of dollars for SMB lead lists · @codyschneider Stripe built an AI platform to boost productivity across all teams · @emilygsands AI tools reduce startup sourcing research to a few prompts · @jongall45 Open autoresearch repo uses LLMs to deeply optimize any domain · @0xSero
Frontier breakthroughs and efficiency gains arrive together
A model reportedly solved 10 open problems in mathematics and theoretical computer science while separate releases showed small models matching flagship performance at a fraction of the cost. One counterpoint: economic complexity may dampen real-world AI takeoff even as raw capability rises.
- OCR gains point to post-training and evals as unexpected levers
- Lower cost-per-inference makes new categories of AI businesses viable
- Individual intelligence gains may not translate to economic acceleration
OpenAI's Astra solves 10 open problems in math and computer science · @polynoamial OpenAI's Astra solved 10 open problems in math and computer science · @polynoamial Economic complexity may slow AI takeoff more than expected · @scottastevenson What changed to make Gemini 2.5 Flash so strong at OCR · @_ontologic Great point · @scottastevenson Sudden OCR improvements usually come from post-training and evals · @_ontologic Inkling-Small matches flagship model at a quarter of the size · @miramurati Luna and Terra's cost/performance makes entire businesses viable · @chrismaddern
Early-stage VC trust and alignment under structural stress
A high-profile fund manager addressed LP rumors of collapse while a parallel critique gained traction that early-stage venture has structurally drifted toward paper-value optimization over genuine company-building. Together they surface a widening gap between how funds are marked and how companies actually perform.
- Series A benchmarks from YC are now explicit and public
- LPs increasingly weight current hustle over stale domain credentials
- The feedback loop between fund marks and real value is blurring
The Series A metrics that separate funded from passed startups · @YCInsight The Series A after YC, what it actually takes ♻️ · @YCInsight This is advice we give our GPs often. · @johnfelix123 Leopold Aschenbrenner addresses LPs after fund collapse rumors · @tbpn Full interview: what happened to Leopold Aschenbrenner · @notthreadguy Cendana Capital values hustle and curiosity over stale credentials · @tbpn Ten emerging VCs from Q2 2026 include deep tech and hard tech funds · @pavelprata Early-stage VC optimizes for stock value, not real company value · @arian_ghashghai Ten emerging VCs from Q2 2026 include deep tech and hard tech funds · @pavelprata Here is the full letter Leopold sent to his LPs last night · @tbpn Early-stage VC now optimizes for stock value over real company value · @arian_ghashghai New report series tracks and analyzes global VC fund closes · @pavelprata
Agentic workflows are crossing from demo to production
Self-improving agent loops have moved from experimentation to standard practice, with production deployments now handling customer support, video editing, and cross-tool work delivery. The question is shifting from whether agents can do the work to who captures the value when they reliably can.
- Fully autonomous customer support and video production are live, not theoretical.
- Recursive self-improvement in agent harnesses is an active systems engineering problem.
- Outcome-based pricing for agents remains the unsettled business model question.
Agent transition favors AI labs while new bundled super apps emerge · @signulll Claude automates 4K video editing and social posting workflow · @brookejlacey Anthropic engineer demos building self-improving agentic workflows · @0xMovez Open dev group building recursive self-improving agent harness · @martin_casado Subagents proposed to audit rollout history and suggest repo docs · @_lopopolo Cloud VMs let you run unlimited AI coding agents in parallel · @ryancarson OpenWorker open-source agent delivers finished work across tools · @AndrewYNg An AI agent now handles all of Gumroad's customer support · @Gumclaw OpenWorker connects your AI model of choice to everyday work tools · openworker.com Agent aims to automate the 75% of work that doesn't need humans · @dotta
The AI inference stack commoditizes layer by layer
Each layer of the AI stack is being squeezed from above: routing cuts inference costs while commoditizing model providers, and value migrates toward vertically integrated experiences. Speed, not throughput, is emerging as the differentiated moat in inference.
- Real-time inference pricing markets are launching, threatening provider lock-in.
- AI labs are unlikely to dominate the application layer despite infrastructure advantages.
- Cheap software shifts rents toward vertical integrators, away from horizontal tooling.
Speed is the key moat in AI inference, not throughput optimization · @firesidealpha Omnious launches real-time inference pricing market for AI models · @Omniousai Cheap software shifts value from SaaS stacks to vertical integration · @naval Cursor launches intelligent router that cuts model costs by 60% · @ccatalini AI labs are unlikely to dominate the application layer, VC argues · @gokulr
Stadium startup culture recruits the wrong founders
The bottleneck on startup progress is converting the right people into founders, not making the path look attractive. Prestige-driven stadium culture and hype incentives recruit status-seekers rather than people willing to endure a decade of demanding work.
- Founding is years of demanding work that stadium framing actively obscures.
- SF's talent crisis deepens as founders' engineers defect to frontier AI labs.
- The real constraint is identifying determined individuals with great ideas.
The real startup bottleneck is turning talented people into founders · @credistick Stadium startup culture may push wrong people into founding · @scottastevenson Startup School moving to stadiums betrays YC's hacker-first identity · @scottastevenson SF founders face existential crisis as AI disrupts talent and meaning · @HugoAmsellem
Power law leaves most venture investors behind
Fewer than 40% of VCs achieve any successful investments, with 5% capturing 90% of profits, and human capital rather than deal skill drives most of the variation. Capital is concentrating further as sophisticated LPs move in, compressing early-stage options.
- Most concentrated portfolio strategies lack analytical or LP-mandated justification.
- IPOs have been a net-negative bet for seven years.
- Former operators entering as LPs are pressuring myopic allocators out.
Top 5% of VCs capture 90% of industry profits, research shows · @arcticinstincts Human capital predicts VC career success and deal access · @nberpubs What makes a VC above average: data, analysis, or value add · @vc Capital concentration and savvy LPs are reshaping early-stage VC · @credistick Many VCs chase concentrated portfolios without real justification · @pavelprata IPOs have quietly been a losing bet since 2019 · @fintechfrank
Open-weight AI threatens closed-model economics and governance
Kimi K3 forced a public reckoning over whether open-weight models undercut frontier AI businesses or simply expose how restrictive and overpriced closed models have become. The debate split along lines of unit economics, safety, and whether soft regulatory pressure could quietly suppress open-weight adoption.
- Closed-model safeguards increasingly read as competitive moats, not safety measures.
- Soft-law agency rules could choke open weights without a congressional ban.
- Linux and the internet precedent suggests openness wins over governability long-term.
New hire walks back Kimi take, clarifies open-weight AI views · @deanwball Open weight AI doesn't slow progress, it breaks closed-model economics · @lior_eth Open-weight AI will succeed for same reasons Linux and the internet did · @ylecun Soft-law strategy could quietly choke off open-weight AI adoption · @WillManidis Kimi K3 exposes how restricted US frontier AI models have become · @DavidVorick OpenAI strategist labels open-source AI communist and decelerationist · @DrNickA Open-weight AI will win for the same reason Linux and the internet did · @ylecun
Eval discipline is the real unlock for production AI agents
A coherent methodology for AI evals emerged across multiple practitioners: start with vibe checks, harden with handwritten scenarios, then close the loop by feeding live production traces back into scoring and prompt optimization. Personal task-specific evals were argued to outperform generic industry benchmarks for day-to-day utility.
- Trace-score-optimize loops can replace expensive offline eval bootcamps.
- Generic benchmarks miss individual capability boundaries that only personal testing reveals.
- Auditing AI decisions, not raw code output, is the key discipline at scale.
Pydantic AI agents self-improve via trace, score and optimize loops · @h100envy Build a personal AI eval set tuned to your own work and tasks · @zarazhangrui Auditing AI choices, not code, keeps codebases from chaos · @VictorTaelin Three-stage framework for building reliable AI evals in production · @ankrgyl Personal AI evals built from your own tasks beat generic benchmarks · @zarazhangrui Reliable evals emerge in a sequence, not all at once · @ankrgyl Audit the decisions AI made, not every line of code it wrote · '@VictorTaelin'
AI adoption is tiny despite overwhelming builder-bubble perception
Household-level data reveals that paid AI usage remains well under 3 percent of the US population, making current builder enthusiasm a narrow elite phenomenon. The gap between Tech Twitter saturation and real-world penetration is large enough to matter for market-size assumptions.
- Only 0.2 percent of US households spend more than 100 dollars monthly on AI.
- Being an active AI builder puts someone in the top one percent of the population.
- Application-layer value accrual thesis depends on mass adoption that has not arrived.
Only 2.2% of U.S. households pay for an AI subscription · @keean_edward AI builders are a tiny minority despite how crowded Tech Twitter feels · '@keean_edward' Surveys show AI adoption remains tiny fraction of US households · @itsolelehmann LLM value may concentrate at the application layer long term · @scottastevenson LLM value will accrue at the application layer, not the foundation · @scottastevenson
Open-source infrastructure commoditizes SaaS dev-tool categories
Self-hosted, zero-marginal-cost alternatives launched this week across email delivery, SEO analytics, and full application deployment, each explicitly targeting the pricing and lock-in of established SaaS incumbents. The pattern suggests a wave of open-source platforms bundling what previously required multiple paid services.
- Transactional email, SEO tooling, and app deployment all saw open-source challengers launch.
- MCP integration is already being bundled into self-hosted platforms for agent workflows.
- Spite-driven development against expensive incumbents is now a stated founding motivation.
OpenShip offers self-hosted email at a fraction of SaaS costs · @openshipio OpenSEO launches as open-source rival to Semrush and Ahrefs · @bensenescu OpenShip launches open-source platform for self-hosted app deployment · @openshipio Open-source SEO tools rise as search shifts toward AI discovery · @bensenescu Open-source platforms now bundle deploy, data, agents and ops in one · @openshipio OpenShip makes transactional email a self-hosted, zero-cost primitive · @openshipio
Agentic infrastructure is now an engineering discipline
Personal and team agents are moving past simple prompting into layered operating systems with routing logic, skill libraries, voice calibration, and self-cleaning tasks. The bottleneck is no longer model quality but the engineering required to make models run smoothly without constant human intervention.
- User-specific attribution in shared agents is an unsolved coordination problem.
- Style and clarity rules embedded in agent configs measurably improve output quality.
- Internal company agents automating 90 percent of busywork signal a new workflow baseline.
For people using codex to draft emails and slacks, try this simple 1x… · @strimblez A well structured personal agent should, work through iMessage… · @ianlapham I'm just going to dump my whole agentic setup out here, because I see… · @jamonholmgren Fable and Sol tend to use a lot of jargon that's hard to understand… · @johnnyheo A small style rule can materially improve agent performance: ask for plain user-facing explanations without sacrificing · @johnnyheo Strong agentic coding is not just better prompting · @jamonholmgren One problem with multiplayer AI is that you want to share agents, but… · @dotta Sierra's internal AI agent automates 90% of coding and busywork · @vijayiyengar The real shift is not just having coding copilots · @vijayiyengar
Foundation models unlikely to win every vertical application
Base model providers can threaten many application categories, but the hard advantages in any vertical lie in workflows, go-to-market, and servicing rather than code, making it unlikely that any single platform company wins across the board.
- Data network effects may still favor large shared models over personalized alternatives.
- Many distinct AI systems trained on local values may outperform one universal model.
- Open source AI infrastructure maps cleanly onto DeFi roles, clarifying competitive dynamics.
Every AI provider that isn't winning the data network effect is… · @scottastevenson It is highly unlikely that the foundation companies will be… · @gerstenzang The strongest AI systems may keep improving because they absorb learning from massive shared user bases, creating a · '@scottastevenson' Base model companies may be able to threaten many application categories, but that does not mean they will win · '@gerstenzang' A DeFi guide to Open Source AI market structure Harnesses are Wallets… · @tarunchitra A useful way to understand open source AI infrastructure is to map it onto DeFi roles · Tarun Chitra Today we share the worldview behind our mission. · @miramurati A good AI future may require many distinct systems rather than one universal model · '@miramurati'
Investor consensus systematically lags founder and market reality
Consensus thinking among investors tends to be a lagging abstraction, rewarding familiar archetypes over genuine outliers and missing the companies that will matter most. The structural incentive inside most venture firms reinforces conformity rather than the ambitious contrarianism they claim to practice.
- Most venture firms depend entirely on one or two partners and will not outlast them.
- Founder-centricity rhetoric often masks a strong preference for a narrow, comfortable archetype.
- LLMs now make LP lookthrough portfolio analysis fast, raising the bar for fund differentiation.
Investor consensus is wrong almost 100% of the time. · @juliey4 This is not to be confused with venture capital’s often-performative… · @credistick Generic investor consensus is often a lagging abstraction built from pattern-matching across markets, while founders · '@juliey4' Venture often claims to be founder-centric while rewarding conformity, polish, and scale theatre over real · '@credistick' Most venture firms behave more like partner-dependent cashflow businesses than durable institutions · @auren 99% of venture capital firms are just lifestyle businesses. · @auren I'm an LP in roughly 25 VC funds. · @andrewparker A lot of venture value comes from small promises kept, not just advice given · '@jgreze' LLMs make fund lookthrough analysis far easier for LPs by turning a tedious manual mapping exercise into a quick way to · Andrew Parker
Current AI cost structures are a temporary inefficiency
Spending heavily on inference for lightweight applications today is analogous to paying high gas fees to mint NFTs in 2021, a phase that will be resolved by optimized cost layers rather than sustained as a permanent equilibrium.
- AI model recommendations are already generating measurable organic product signups.
- Investors buying irreplaceable assets like sports franchises suggests a hedge against AI commodification.
- Years in crypto are no longer a reliable proxy for founder quality as tooling matures.
Spending $1000 on inference over the weekend to vibecode your toy app… · @johnpalmer In April @ThriveCapital launched a permanent capital vehicle for… · @ssokol94 An observation The number of years in crypto used to be strong signal… · @CarlKVogel Some of the investors funding AI are redirecting gains into assets technology cannot easily commodify, such as sports · '@ssokol94' Spending heavily on inference for lightweight AI apps is a temporary phase, not a durable equilibrium · John Palmer Years spent inside crypto used to be a strong proxy for founder quality because the ecosystem was technically obscure · '@CarlKVogel' AI model recommendations are driving a new wave of organic signups · @zenorocha Some products are starting to get meaningful organic growth because language models recommend them directly to users · @zenorocha
Agents reshape dev workflows but not the need for comprehension
AI coding agents are replacing traditional specs with leaner formats and automating codebase discovery, yet a strong counter-current holds that developers who cannot read and understand agent-generated code will lose the ability to catch errors and maintain systems.
- Product specs are replacing PRDs as the primary unit of product work
- Internal AI tools now handle both execution and institutional memory at fast-growing companies
- Scanning for unknown unknowns in a codebase is becoming a standard workflow step
My take · Jess · Jul 4
While many fear what ai will mean for startups and software products, I'm a big believer that the firm is becoming software, and startups have an advantage at first principles rethinking how entire industries are serviced with ai.
New coding workflow probes known and unknown gaps in codebases · @dzhng Sierra says internal AI agents became core to running the company · @noahhorner_ Product specs replace PRDs in agent driven software teams · @gokulr Developers still need to understand code written by agents · @kieranklaassen Agent loop turns user interviews into paying software · @pzakin Developers still need to understand code written by agents · @geoffreylitt
Frontier AI moat claims are fragmenting under pressure
The winner-take-all narrative around frontier models is under strain as open-source alternatives close capability gaps and enterprise trust in closed labs erodes, while competing moat claims across model providers, harness builders, and token sellers undercut any single dominant position.
- Benchmark gains on weaker models do not replicate frontier output quality
- Open-source hosting lets enterprises sidestep data-sharing concerns entirely
- Cost pressure from open models is forcing a rethink of AI security assumptions
Venture capital built the wrong AI and pension savers may pay · @AndrewOrlowski AI vendors push conflicting moats across models, tooling and tokens · @0interestrates Benchmark gains do not prove weaker models match frontier AI · @alexframegreen Claims that AI will be winner take all may not hold · @sebkrier Open models gain ground as trust in frontier labs frays · @gbrl_dick AI capability gains still look early, not near commoditization · @deanwball Open source model costs are forcing enterprises to rethink AI security · @quxiaoyin AI progress may depend on economic capability, not tokens · @lefttailguy
Venture pricing detaches from fundamentals, echoing 2021
Senior partners at prominent funds are warning that valuations have become untethered from reality in a pattern matching 2021, while megafund fee structures, SPV fraud in secondaries, and geographic bias compound the disadvantage for limited partners and non-Bay Area founders.
- Angels function as credibility signals and intro nodes, not only capital sources
- Megafund fee drag can eliminate hundreds of millions in LP returns versus PE peers
- Secondary market fraud and ghost shares go underreported because victims find it embarrassing
Angel investing works on belief and affinity, not VC logic · @jayyeh Early angel rounds work best when founders start close to home · @christophersaum Top venture investors warn the funding boom looks like 2021 · @madhavchanchani VC megafund fee terms can erase gains for LPs · @credistick Founders should build an angel base before pitching VCs · @jayyeh Fraud and ghost shares taint startup secondary investing · @rsg European founders struggle to raise from Bay Area VCs · @arian_ghashghai
Token and equity alignment stays structurally unresolved in crypto
As projects raise equity alongside tokens, the question of whether token holders and shareholders have aligned incentives remains contested, with protocol-level burn mechanics earning the most trust and dual structures drawing sustained skepticism.
- Long-duration warrants on tokens signal genuine institutional conviction
- Robinhood's DeFi yield mixes native lending with incentive campaigns on a rival's rails
- Utility tokens with on-chain revenue visibility offer a rare verifiable case
Robinhood 7 percent DeFi yield relies heavily on incentives · @HorinekPM Robinhood DeFi launch ties a rival product to Coinbase rails · @sealaunch_ Venice token thesis leans on burns, profits and agent markets · @nikshepsvn Venice offers a rare case for a true AI utility token · @DrNickA Crypto token structures hinge on trust and alignment · @knimkar Crypto projects may still need both tokens and equity · @lex_node
Enterprise AI trust and security concerns grow
Anthropic's developer tools face escalating enterprise trust problems, with unverified spyware allegations and claimed geo-targeted surveillance checks in Claude Code driving bans, while questions about whether AI outputs can be reliably verified at all deepen the credibility gap.
- Alleged timezone checks targeting Chinese users have prompted enterprise-wide bans
- Fable 5 jailbreaks reportedly added no novel capabilities beyond existing models
- Hard-to-verify AI output signals a product design failure, not just a technical limit
My take · Jess · Jul 3
While interest in this seems to be picking up, Im surprised by how little concern enterprises have over protecting their data and workflows from model providers or integrators. I expect this trend of increased caution to continue.
Anthropic restores Fable 5 with tighter cyber safeguards · @alexstamos Security concerns trigger enterprise ban on Claude Code · @__tinygrad__ Alleged client-side checks raise trust concerns around coding tools · @AISuperDomain Poor AI product design can make outputs impossible to verify · @HamelHusain
AI economic value migrates below the model layer
The competitive question has shifted from model capability to where margin actually concentrates. Workflow orchestration, enterprise flywheels, and edge distribution are all contending as the real prize, while model providers face commoditization pressure from multiple directions.
- Embedded workflow lock-in may outlast any model-level advantage
- Enterprise flywheels capturing tacit knowledge could prove more durable than model leads
- Edge providers capturing economic value have strong incentive to keep it private
Applied AI value may accrue to workflow orchestration layers · @i_zerbib Applied AI value may accrue to workflow orchestration layers · @levie Enterprise AI moats may come from embedded workflow lock in · @alexolegimas Enterprise AI shifts toward proprietary model flywheels · @AravSrinivas AI stack specialization could mirror cloud's giant winners · @masonnystrom Frontier AI labs face pricing pressure at the edge · @hypersoren Enterprise AI may spread faster through chat than seat licenses · @_simonsmith AI startup moats are diverging across labs data and distribution · @aashaysanghvi_
Frontier model access shifts to government approval gates
Government approval gates emerged for GPT-5.6 access on security grounds, as a massive distillation attack on Claude attributed to Alibaba confirmed that capability extraction is a live threat. Federal AI policy simultaneously hardened toward restricting Chinese frontier model access globally.
- Gated rollout creates a two-tier market of approved and non-approved users
- Distillation attacks may become the primary vector for capability transfer between rivals
- Regulatory trajectory mirrors the protocol-level battles crypto already fought
Security restrictions could shut Chinese frontier models out globally. · @teortaxesTex Federal AI policy shifts toward tighter and murkier control · @deanwball OpenAI unveils GPT-5.6 lineup with three model tiers · @OpenAI AI labs face the same regulatory battles crypto already fought · @ccatalini U.S. moves to approve GPT-5.6 access customer by customer · @ns123abc OpenAI asked to stagger GPT-5.6 launch over security fears · @steph_palazzolo Anthropic says Claude faced a massive model distillation attack · @MTSlive
Agents split enterprise AI into two distinct problems
Building with agents divides into two fundamentally different challenges: restructuring internal operations versus rebuilding products as agents, and conflating the two is a key failure mode. Internally, agent adoption is already visibly compressing white-collar roles and dissolving traditional team structures at leading firms.
- Role boundaries between engineering, product, and design are dissolving
- Sharing successful agentic workflows across a team remains unsolved
- Executives treating AI as a compliance checkbox are misreading the transition
Agent adoption splits between internal operations and products · @ankrgyl AI teams may reorganize around product lifecycle archetypes · @bcherny Executives warned AI will remake companies, not tick a box · @Steve_Yegge Teams still lack good ways to share agent workflows · @dwr Teams still lack good ways to share agent workflows · @dwr Agentic workflows could compress white collar labor demand · @_The_Prophet__ Meta product teams replace process with internal AI agents · @nikhyl
Moats thin as mega fund capital distorts every stage
Conventional defensibility is weakening to where founder reputation has become the primary moat signal, with teams outweighing decks in early-stage evaluation. Mega fund over-capitalization is simultaneously lowering bars across all stages, inflating rounds throughout the funding stack rather than specifically targeting seed.
- Sitting out a bubble may carry more career risk than joining it
- Seed fund strategy is shifting toward diversification and option value
- A large liquidity wave may be approaching across major private companies
AI startup moats may come from multiplayer systems, not models alone · @illscience Startup moats weaken as investors fixate on teams · @zamdoteth Seed funds shift toward broad portfolios and option value · @masonnystrom Venture firms brace for a new wave of startup liquidity · @Samirkaji New research examines how active mega funds are in seed · @pavelprata Venture incentives can reward joining bubbles over sitting out · @TurnerNovak Mega funds may be inflating startup rounds across every stage · @gdibner Venture capital may split between swarms and permanent capital · @joeljohn
Open models are closing the frontier performance gap
As capable open models approach near-parity with closed frontier models, the cost differential is becoming decisive. Frontier labs have priced for margin rather than compute cost alone, and that margin looks fragile when open alternatives offer roughly 90% lower cost at a 10% performance penalty.
- Subsidized inference may have been masking open model viability all along.
- Token cost pressure at scale is pushing legal AI toward post-training open models.
- Capital allocated to frontier AI could be mispriced if intelligence commoditizes.
Frontier model pricing reflects quality, margin and scale economics · @tugot17 Harvey turns to open models as token costs crush margins · @willchen500 GLM 5.2 could ignite a lasting boom in rented GPU demand · @goodalexander AI teams face a brutal new era of token ROI and cost control · @adambcohen93 GLM-5.2 could be the first open model to replace closed ones · @gneubig AI capex may overshoot demand and trigger a boom-bust cycle · @orrdavid Open source AI progress could push governments toward censorship · @goodalexander A 10% intelligence gap at 90% lower cost with $5 trillion at stake · @zerohedge The chart that breaks the AI trade when intelligence gets cheap · @_The_Prophet__
Multi-agent orchestration reframes what a model is
The Sakana Fugu launch illustrates a structural shift where frontier-level performance is achieved by orchestrating swarms of smaller models rather than scaling a single one. Intelligence is beginning to look less like a product and more like a supply chain.
- Codex-style loops can audit, test, and fix entire codebases autonomously at scale.
- Effective loops require real-browser verification and environment tooling, not just prompts.
- The single-model API abstraction increasingly conceals a multi-agent system beneath.
Codex loops can audit, test, and fix apps at massive scale · @tomosman Builders are swapping examples of Codex and Claude loops · @argofowl Sakana launches Fugu to match frontier AI through orchestration · @SakanaAILabs AI orchestration turns models into a commodity supply chain · @_The_Prophet__ Agent loops get stronger with better testing and tooling · @ctatedev Sakana Fugu runs a learned multi-agent swarm behind one model API · sakana.ai A builder names the three-tool AI agent stack as his best workflow · @yitong
Venture capital concentration is distorting deal quality
Capital has consolidated into larger funds whose fee structures push toward consensus bets and narrative-aligned investments, leaving genuine outliers chronically underfunded. Seed is effectively two markets, with roughly 10% of deals capturing half the dollars and most headlines.
- Emerging managers lost ground during the ZIRP boom, not after it.
- Fee incentives drive GPs toward scale and safety rather than discovery.
- Founders can resist valuation pressure by forcing investors to name a counter-number.
Founders should force investors to make real pricing offers. · @amiklas Fund economics push venture investors toward safer consensus bets · @gdibner A small hype segment dominates seed capital and media attention. · @tmrohan Venture is consolidating into larger funds over diversification · @credistick Capital concentration has distorted venture more than most admit · @Rick_Zullo Early-stage venture is having a standout year as a category · @adityaag Emerging managers lost ground through ZIRP, not the reverse · @credistick
AI is splitting engineering teams and repricing software
Generative coding tools are fracturing engineering into those who generate and those who review, creating a visible class divide and a profession-wide identity crisis. Simultaneously, software pricing is shifting away from seat licenses toward alignment with actual business outcomes.
- The 'context layer,' infrastructure making AI useful for real codebases, is a major emerging VC thesis.
- Coding agents remain too 'software-brained' to generalize cleanly to broader knowledge work.
- Taking humans fully out of high-stakes loops remains a 'brutally long slog' even with mature tooling.
Coding agents still struggle to fit real knowledge work · @emollick Startups race to build the context layer for engineering AI · @GergelyOrosz AI coding is splitting engineering teams into pushers and reviewers · @deedydas AI is pushing software from seat sales toward paid business outcomes · @rodriscoll Taking humans out of the loop is brutally hard but worth it. · @Jason Howie hit 50% autopilot after years of human-in-loop scheduling · @awwstn
Intent-centric OS and data firewalls reshape platform power
Apple's announced shift toward intent-centric interfaces that construct UI on demand threatens the app ecosystem model. Platforms blocking LLM data access are spawning decentralized crawling marketplaces as a countermeasure.
- On-demand interface generation could eliminate third-party apps entirely
- Permissionless data markets emerge wherever platforms erect LLM firewalls
AI agent costs exploded, forcing outcome-based pricing models
The shift from chat to agents drove costs far beyond early estimates, as recursive agent spawning multiplied token consumption. Firms that price for outcomes over token volume can capture labor-scale revenue, but only by treating every token and hour as margin.
- Token spend at scale blurs the line between software and services
- Open question: which firms can actually measure outcome quality reliably
Real autonomy is finding work, not executing tasks
The defining feature of a truly autonomous loop is not task execution but the ability to propose work without a human prompt. Self-improving systems that search proactively over potential improvements against stated objectives represent the next capability threshold.
- Proactive explorer agents differ fundamentally from reactive task agents
- Self-improvement applies at the organizational level, not just software
Individual AI assistants give way to team orchestration
The market needs an orchestration layer spanning an entire team's workflow, combining work management, agent assignment, and shared context. Codified, programmable workflows are replacing knowledge-based skills as the reliable unit of agent execution.
- Shared MCP servers and context compound value across teams
- Codified workflows require less model intelligence, enabling more reliable execution