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 collapses timelines from weeks to hours 6 signals ▾
Founders can compress weeks of market research into hours with Claude · @fin465 MCP-native fundraising turns capital raises into agent workflows · @harris The interesting shift is not agents summarizing papers · @transformerlab YC method compresses a month of GTM research into 3 hours with AI · @fin465 Primus AI agent automates the full ML research loop · @transformerlab Free MCP-powered fundraising platform launches in beta · @harris
AI is eliminating the time cost of knowledge-intensive processes across domains. GTM research that took a month compresses to hours by feeding structured primary sources, fundraising becomes a fully agent-driven workflow, and autonomous systems now run the complete scientific loop without human intervention.
- Structured primary sources beat open-ended 'research this market' prompts
- Autonomous research agents hypothesize, experiment, and write papers independently
Kingmaking concentrates capital away from venture's discovery mission 4 signals ▾
Seed portfolio construction is not just about valuation discipline · @nchirls Seed GPs balance non-consensus and consensus-priced founder bets · @nchirls VC kingmaking narrative conflicts with the industry's founding mission · @arian_ghashghai Venture kingmaking turns VCs into private market stock traders · @arian_ghashghai
The 'only 20 companies matter' narrative effectively turns venture capital into private market stock trading, prioritizing known winners over the discovery of non-consensus founders. Seed portfolio construction still requires the discipline to back non-obvious founders before they become obvious ones.
- Overrigid ownership targets create adverse selection at the seed stage
- LPs need to trust GPs to pursue non-consensus bets, not just consensus ones
via @nchirls, @arian_ghashghai
Nvidia underwrites its own GPU depreciation risk 2 signals ▾
Banks have resisted financing GPU purchases because depreciation is unpredictable, driven by Nvidia's own product roadmap. Nvidia now offers depreciation insurance directly to lenders, completing a loop where the company that creates obsolescence also insures against it.
via @8teAPi
Mature agent stacks need harnesses, not just models 10 signals ▾
An agent harness is more than a model wrapped with tools · @RubricLabs Giving agents direct access to passwords is the wrong architecture · @0xZoZoZo What an agent harness is · @RubricLabs AI agents get secure credential access via 1Password service accounts · @0xZoZoZo Logistics AI agents learn new skills autonomously via memory graphs · @typesfast Used-car platform runs 95% of transactions on AI agents · @VirtualElena Remoko lets AI agents send iOS push notifications via MCP · @yoheinakajima Remoko lets AI agents reach you with iOS push notifications · @yoheinakajima Logistics AI agents now learn new skills autonomously via memory graphs · @typesfast The real edge in AI-native companies is not just adopting agents · '@VirtualElena'
Production agents require more than a model wrapped with tools. Emerging architecture separates credential management into isolated vaults, adds memory graphs for autonomous skill accumulation, and demands that organizations restructure around persistent agents rather than simply adopting them.
- Agents need isolated credential vaults, not direct password access
- Memory graphs let agents accumulate new skills without human instruction
- Organizations should self-improve with each model release, not just individuals
Latest signals · Tuesday, August 11 all signals →
new An agent harness is more than a model wrapped with tools · @RubricLabs new Seed portfolio construction is not just about valuation discipline · @nchirls new Giving agents direct access to passwords is the wrong architecture · @0xZoZoZo new Nvidia converts roadmap knowledge into depreciation insurance for banks · @8teAPi new Zuckerberg wants his values known before stronger models arrive · @alexeheath new MCP-native fundraising turns capital raises into agent workflows · @harris new The interesting shift is not agents summarizing papers · @transformerlab new Founders can compress weeks of market research into hours with Claude · @fin465 Earlier weeks
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
Agentic execution scales faster than verification or human adoption
AI output has expanded dramatically while enterprise adoption and verification tooling lag significantly behind. The asymmetry between cheap agentic execution and costly outcome validation is becoming the defining friction of the current deployment moment.
- Large org deployment requires navigating seven layers of internal process
- Execution cost falls fast while claim verification cost barely moves
- Recursive self-improvement tooling may signal actual exponential takeoff has begun
AI output is exploding while human adoption stays flat · @jenzhuscott Even working AI faces seven gates of operational hell in big orgs · Brandon Carl AI execution is getting cheap faster than verification can keep up · cgi.md Recursive self-improvement tools mark early exponential liftoff · Elad Gil The last six months may be among the most important in history · Elad Gil
One operator with parallel agents achieves team-scale output
A single person orchestrating 20 to 30 parallel agents can now match the throughput of engineering teams, investment analysts, and content operations. The emerging unit is one orchestrator setting objectives and reallocating compute, not a traditional team structure.
- One engineer shipped 40 pull requests a day using parallel agents
- In YC's spring batch, 60% of one-liners mention AI or agents
- The org chart flattens as compute replaces headcount scaling
Solo engineer ships 40 PRs a day running 20-30 parallel agents · Peter Yang (featuring Kun Cheng) A solo investor replaces a whole team with AI agent workflows · @zeroxkyle Naval's rule for the AI era is waste tokens to save human time · Naval Founders will run teams of agents like portfolio managers setting OKRs · deana Three-agent content pipeline cuts post production from 30 hours to 4 · Payton Kaleiwahea (summarizing Alex Lieberman) YC Spring '26 is the agents batch, 60% pitch AI in their one-liners · Ollie Forsyth Distill frontier models into local agents for a private personal assistant · Tomasz Tunguz
Open models match closed ones, pricing gap stays wide
The capability gap between open-weight and closed models has closed faster than expected while pricing has barely moved, creating immediate arbitrage for builders who route across providers. Application vendors that stay provider-neutral and charge on outcomes rather than tokens are finding structural cost advantages.
- Lindy's switch to DeepSeek cut costs and improved performance simultaneously
- Model routing becomes a primary lever for cost and risk management
- Charging on outcomes rather than inference tokens realigns incentive structures
Open and closed model capability converged, but pricing didn't · Chamath Palihapitiya AI apps beat the labs by staying Switzerland and routing across every model · Brad Menezes Lindy moves all traffic off Anthropic to DeepSeek v4, cheaper and better · Flo Crivello Raw intelligence will get cheap, but outcomes don't compress · Sierra AI conventional wisdom keeps reversing, from wrappers to desktop GUIs · Greg Isenberg OpenAI and Anthropic will move up the stack and absorb Cursor and Lovable · Jason Calacanis
Venture incentives misalign as fees rise and returns fall
Top VC funds now extract fees at many times their historical rates while returns have weakened relative to public markets. Early-stage culture has shifted toward performative input metrics, rewarding token burn and launch visibility over genuine product building.
- The venture power law at fund level is largely self-inflicted
- Benchmark's new growth fund signals a structural shift in firm strategy
- AI-native company speed is outpacing traditional slow-moving venture processes
Benchmark raises $2B and launches its first growth fund · Kate Clark AI investing hits cost walls as power-law belief breeds concentration · Brandon Carl Traditional VC's slow process clashes with AI-native company speed · Dave Fontenot Silicon Valley traded weirdos and tinkerers for input-maxxing strivers · Andrew Ziperski The venture power law is self-inflicted at the fund level · Dan Gray (Odin) Early-stage venture now rewards performative input-maximizing over building · Dave Fontenot Striver founders input-max performative metrics over real product fit · Andrew Ziperski Top VC firms now extract 10x more fees as their returns weaken · Dan Gray
Deterministic orchestration outperforms emergent multi-agent coordination
Structured, explicit workflows driving small agent loops are proving more reliable than non-deterministic multi-agent systems for complex tasks. New platform features and DSLs with primitives like parallel() and pipeline() are making this architecture the practical standard.
- Explicit workflow primitives replace emergent coordination between autonomous agents
- Structured orchestration unlocks multi-step tasks previously too complex for agents
- Context engineering, not model capability, is the remaining differentiator
Deterministic workflows orchestrating small agent loops beat agent soup every time · dex (@dexhorthy) New DSL gives subagents explicit workflow code instead of emergent coordination · Michael Livs (@micLivs) The /workflow feature cracks multi-step tasks agents long failed at · Jason Zhou Deterministic workflows around small agent loops beat multi-agent soup · dex Claude Code auto-generates orchestration plans it follows across hundreds of agents · cat
Infrastructure, not model quality, is now the real constraint
Model capability has crossed a threshold where harness, connectors, and reliable uptime matter more than raw intelligence. With nonhuman traffic exceeding half of some APIs, the build question has shifted from 'can the model do this' to 'can the system stay up and integrate.'
- No unified control plane yet exists for multi-domain agent orchestration
- Products lacking headless interfaces risk repeating the early mobile-era mistake
- Agents are beginning to manage their own threads, worktrees, and orchestration
Agents now need only a harness, connectors and reliable uptime, not better models · @nickbaumann_ No single control plane exists to manage agents across code, design and ops · Dennison Bertram Codex now manages its own threads and spins up worktrees for parallel tasks · Guinness Chen Agents are the next mobile, with 53% of one API's traffic already nonhuman · Gokul Rajaram Shopify's River is an AI agent living inside company Slack · shopify.engineering Sepo agent runs in GitHub Actions to implement issues and review PRs · github.com
Professional networks are becoming AI-queryable deal infrastructure
A set of tools now treats relationship graphs as structured, searchable data for finding warm backchannels, sourcing founders, and activating sales. The workflow collapses what previously required hours of manual research into a single natural-language query.
- Warm intros are closing more stuck deals than cold outreach
- Early founder signals surface from follower graphs before pitches arrive
- Network search now reaches across teammates' and friends' connections
Agents complement humans rather than replace them near-term
Spending on agents shows sharply diminishing returns beyond a baseline, while the competitive split is between people using AI and those who are not. The optimism and judgment required to act under uncertainty remain human advantages.
- Returns to agent expenditure fall off faster than for human labor
- The competitive divide is humans with AI versus humans without
- Intense conviction about outcomes remains a human edge agents cannot replicate