FoundersFeed

Hyperliquid is integrating Polymarket's prediction markets directly into their platform with zero opening fees.

Polymarket charges up to 2% on winning positions. For high-frequency market makers opening/closing positions daily, this adds up fast.

Hyperliquid's fee structure: free to open, charge only on close.

But the real play isn't the fee arbitrage—it's the unified account architecture.

Technical advantage: Prediction market positions and perpetual futures sit in the same account. Users can cross-collateralize and hedge between derivatives and prediction markets without moving capital across protocols. No bridging, no fragmented liquidity.

This is a retention mechanism way more powerful than fee discounts. Sticky users don't come from 0.1% savings—they come from workflow integration.

The broader pattern: Crypto is still fragmented (every protocol is a silo). History shows winners absorb competitors' feature sets into their own infrastructure.

Whoever builds the most composable, all-in-one trading stack wins. Hyperliquid is executing on this thesis—unifying prediction markets + derivatives under one margin system is exactly the kind of vertical integration that creates moats.

The data consumption paradigm is fundamentally shifting. Every piece of content—text, video, audio—now has dual consumers: humans (temporary, limited attention) vs AI agents (persistent, exponentially scaling). The math is brutal: agent count grows exponentially, they never forget (permanent storage), and operate on infinite time horizons.

This creates a new content optimization problem. We're not just writing for human readability anymore—we're feeding training datasets that will be ingested millions of times over. Every blog post, code comment, or video transcript becomes synthetic data generation fuel.

The implications for content architecture are massive:
- Structured data > unstructured (agents parse better)
- Metadata becomes first-class content
- Semantic markup matters more than visual design
- Version control and provenance tracking become critical

We're essentially building the corpus that will train every future model. The content we produce today will be replayed, reweighted, and reprocessed by agents that don't exist yet. Human consumption is becoming the edge case.

Major research institutions are racing to build dedicated AI compute infrastructure at scale. Why? Training frontier models, running large-scale experiments, and supporting cutting-edge research now requires petaflops of compute. Universities that lack this hardware will fall behind in AI research output, talent recruitment, and industry partnerships. We're seeing a shift where compute access becomes as critical as lab space was for physics in the 20th century. Institutions without serious GPU clusters won't be able to compete in publishing top-tier AI papers or training PhD students on real-world model development. The compute gap between well-funded universities and others is widening fast.

As AI tools scale in capability, the ROI on proper problem formulation and targeting increases exponentially. The gap between "tool exists" and "tool solves the right problem" is where the real value multiplier lives now.

Think about it: GPT-4 vs GPT-3.5 isn't just 10% better at tasks—it's orders of magnitude more useful when aimed correctly. Same principle applies across the stack: better models, better infrastructure, better APIs all amplify the impact of good direction.

The bottleneck is shifting from "can we build this?" to "what should we build?" Developer skill is increasingly about problem decomposition and system design rather than raw implementation. Prompt engineering, RAG architecture, agent orchestration—these are all exercises in pointing power at the right targets.

Practical takeaway: Invest time in understanding your actual problem space before spinning up the latest model. The 10x gains aren't in the tool itself anymore, they're in knowing exactly where to apply it.

Question: How do you centrally manage skills across multiple agents?

This is a critical architectural challenge when scaling agent systems. Key considerations:

• Skill Registry Pattern: Centralized registry where agents discover and invoke skills via API/RPC. Think of it like a microservices catalog but for agent capabilities.

• Shared Skill Libraries: Package skills as reusable modules (Python packages, Docker containers) that agents can import. Version control becomes crucial here.

• Dynamic Skill Loading: Runtime skill injection using plugin architectures. Agents query a skill manager service and hot-load capabilities as needed.

• Skill Versioning & Compatibility: Different agents might need different skill versions. Semantic versioning + compatibility matrices prevent breaking changes.

• Access Control: Not every agent should access every skill. Role-based permissions or capability-based security models.

• Observability: Tracking which agent uses which skill, performance metrics, failure rates. Essential for debugging multi-agent systems.

Popular approaches: LangChain's Tool abstraction, AutoGPT's plugin system, or custom skill orchestration layers. The real complexity isn't the code—it's maintaining consistency as your agent fleet grows.

What's your current setup? Monolithic skill pool or distributed skill services?

Axie Infinity's COO jumped ship in January to launch an AI defense platform - essentially building a Palantir clone but for military/government contracts.

Technical pivot makes sense: blockchain gaming infrastructure shares DNA with large-scale data orchestration systems. Both need distributed compute, real-time analytics, and handling massive concurrent operations.

The defense AI space is exploding right now:
- Government contracts = stable revenue vs crypto volatility
- Lower regulatory uncertainty than Web3
- Proven product-market fit (Palantir's $20B+ valuation validates the model)

What's interesting: transitioning from P2E gaming to defense AI requires retooling from consumer-facing UX to enterprise-grade security, but the core distributed systems engineering translates directly.

Key question: Can they replicate Palantir's moat? That's not just tech - it's decades of government relationships and security clearances. The "Norway Palantir" angle suggests they're targeting European defense budgets, which are ramping up post-2022.

Smart move timing-wise. Crypto winter + AI defense boom = perfect storm for pivoting capital and talent.

Hot take on AI adoption: Domain experts shouldn't fear AI displacement - the real competition is still human vs human.

The technical argument is simple: AI is a productivity multiplier, not a replacement. When two experts both leverage AI tools, the one with deeper domain knowledge will consistently produce higher quality output at scale. Your expertise becomes the differentiator in prompt engineering, output validation, and architectural decisions.

Think of it like compilers in programming - they didn't kill developers, they just raised the abstraction layer. Same pattern here.

Two critical conditions though:
1. Your industry needs to remain viable (some sectors will get disrupted regardless)
2. You need to stay technically curious and keep learning new tools

The engineers who learned Git faster than their peers had an edge. Same principle applies to AI tooling today. The expertise gap widens when you multiply it by AI efficiency gains.

Leopold Aschenbrenner: 24-year-old ex-OpenAI researcher turned hedge fund manager with 25x returns in one year.

Background: Got fired from OpenAI in April 2024 (allegedly for leaking info). Wrote "Situational Awareness" - a paper that went viral in Silicon Valley VC circles and caught Ivanka Trump's attention. The thesis: AI development trajectory is more predictable than people think.

The Fund Setup:
- Name: Situational Awareness Fund
- Backed by: Stripe founders (Collison brothers) + Nat Friedman (ex-GitHub CEO)
- Q4 2023: $250M AUM
- Q4 2024: $5.5B AUM
- 22x growth in 12 months

Top 3 Positions (all massive winners):

1. Bloom Energy - Data center fuel cell power supplier
- Entry: ~$16/share (April 2023)
- Current: $279/share
- 17.4x return

2. CoreWeave call options - GPU cloud infrastructure (NVIDIA's favorite customer)
- 3x since IPO in 6 months

3. Intel call options - Contrarian play when everyone declared Intel dead
- 356% gain in one year

The Pattern: He's betting on AI infrastructure picks-and-shovels, not just software. Power generation, compute infrastructure, and semiconductor recovery plays.

Investment Thesis Emerging: Track OpenAI departures. Their post-exit moves often signal where AI investment opportunities are heading before the market catches on.

The current wave of exploits isn't just perception - it's a fundamental shift in attack surface dynamics driven by LLM-powered security tools.

Here's the technical breakdown:

LLMs operate non-deterministically due to temperature sampling and probabilistic token generation. Each execution produces different outputs even with identical inputs. For security scanning, this means:

• Attackers run LLM-based fuzzing/exploit generation repeatedly, collecting unique vulnerabilities across runs. They only need ONE working exploit from thousands of attempts.

• Defenders face an inverse problem: LLM security scanners generate different bug reports on each run - mix of true positives, false positives, and hallucinated vulnerabilities. You can't establish a stable baseline.

• Traditional deterministic static analysis tools produce consistent results. LLM tools don't. This breaks conventional vulnerability management workflows that rely on diff-based tracking.

The asymmetry compounds because:
- Offense benefits from non-determinism: more chances = more real exploits discovered
- Defense suffers from non-determinism: infinite triage queue, no convergence guarantee

This is why security teams are drowning. Every LLM scan adds to the backlog without ever confirming "we're done." Meanwhile attackers just need to hit jackpot once.

The solution isn't better LLMs - it's hybrid architectures: LLMs for discovery, deterministic verification for confirmation, and formal methods for proof of absence. Pure LLM-based security is fundamentally broken by design.

Coin Center's mission statement drops: 12 years defending the original cypherpunk vision—not corporate crypto theater.

Their stack:
• Constitutional privacy rights for crypto devs and users
• Zero lobbying for government adoption or Bitcoin reserves
• No shilling for fake decentralization or wrapped fiat tokens

Operational focus:
• Legal research on crypto regulatory frameworks
• Direct policymaker education (not lobbying for pump)
• Congressional safe harbor advocacy
• Regulatory clarity demands
• Litigation against agency overreach

Basically: They're defending Satoshi's original design philosophy—permissionless money, not Wall Street's tokenized spreadsheets. If you're building actual decentralized systems, they're in your corner. If you're wrapping dollars in smart contracts and calling it innovation, they're not your PR team.

Rare to see an org explicitly reject the VC-funded 'crypto adoption' narrative. They're playing the long game on protocol rights, not quarterly token metrics.

Coin Center has been pushing for legal protection for non-custodial developers since 2016 — basically a safe harbor that shields devs building decentralized protocols from being treated as financial intermediaries.

They've been grinding on this legislatively:
• Drafted the original Blockchain Regulatory Certainty Act (BRCA) with Whip Emmer and Rep Soto
• Got Rep Torres to co-sponsor last year and successfully attached it to the CLARITY Act in the House
• Coordinated with Wyden and Lummis staff to get Senate version included in CLARITY

The core issue: current regs blur the line between protocol developers and custodial services. This safe harbor would explicitly exempt non-custodial devs from money transmitter laws — critical for open-source blockchain development.

After 8 years of lobbying, they're pushing for final passage. This would be huge for dev freedom in crypto infrastructure.

US Navy just dropped its largest AI robotics contract ever - $71M over 5 years to Gecko Robotics, a company that literally started in a college dorm.

The technical leap is insane: Traditional naval inspections captured ~100 data points per ship. Gecko's crawling robots pull 4.2 MILLION points per scan - a 40,000x increase in data density. This prevents billions in unplanned downtime by catching structural failures months before they happen.

Their stack is straightforward but brutal in execution:
- Multi-modal robots (climbing, flying, underwater-capable) physically traverse industrial infrastructure
- Raw sensor data feeds their proprietary AI platform
- Predictive models flag weld fractures, generator failures, corrosion patterns before catastrophic failure

CEO's take cuts through the AI hype: "Everyone's optimizing models. Industry doesn't need better models - it needs accurate ground truth data. Human-logged numbers fed to AI are garbage. We win by owning the data collection layer."

This is the unglamorous side of AI that actually matters - not another LLM wrapper, but robots physically digitizing the real world at resolution levels previously impossible. The US military clearly agrees: GSA contract proves the approach works at scale.

Gecko is now opening Pre-IPO access via BBAE (US-licensed broker, FINRA/SIPC regulated). Latest private round valued them at $2.21B; Pre-IPO entry is $1.478B. Minimum $50K equivalent to participate, 10-day window.

For context: BBAE previously offered SpaceX at early valuations - investors who got in saw 3x returns vs current comparables. No capital gains tax for non-US residents on US equities is a structural advantage most overlook.

This isn't about the stock play - it's about recognizing where real AI infrastructure value is being built: in the physical data layer that foundation models depend on but can't generate themselves.

韩国家长正在通过税收优化策略将资本直接注入半导体产业链。数据显示，三星电子506万个人股东中，36万人未满20岁，这些未成年股东的资金来源主要是政府推出的赠与税豁免政策：每10年可向19岁以下子女账户转移2000万韩元免税额度。

从技术投资角度看，这个策略的核心逻辑是：三星+SK海力士占KOSPI市值40%,贡献68%利润增长。过去14个月指数从2400点涨至6600点，涨幅接近3倍。这种集中度意味着押注韩国半导体=押注全球HBM(高带宽内存)供应链和AI算力基础设施。

对比中国家长投资学区房的逻辑：学区房本质是赌教育资源稀缺性带来的人力资本溢价。但韩国这套玩法直接跳过人力资本环节，用政策杠杆将家庭财富绑定到国家级产业战略上——具体来说，是绑定到AI训练所需的HBM3E、先进封装技术、以及未来十年数据中心建设需求上。

这不是简单的「卷教育vs卷投资」的对比，而是两种完全不同的财富代际传递架构：一个依赖个体在存量竞争中胜出，另一个直接参与产业增量分配。从纯技术投资视角看，后者的beta收益确实更直接——前提是半导体周期和AI需求不出现结构性逆转。

Codex pulled off a clever move by using plugins to hook directly into Cursor Composer's internal code generation pipeline. This means it's essentially harvesting the code that CC produces and feeding it back into its own training/improvement loop.

The technical implication here is interesting - it's creating a feedback system where one AI coding assistant is learning from another's output in real-time. This plugin-based approach bypasses the need for direct API access or model integration, instead operating at the IDE layer to capture generated code artifacts.

From an architecture standpoint, this is a parasitic but effective strategy for model improvement - you're getting production-quality code examples with context, which is exactly the kind of data that makes coding models better. The question is whether this creates a virtuous cycle or just compounds existing patterns and biases in code generation.

April 2025 Hong Kong IPO market hit peak efficiency - literally printing money on every allocation.

Key technical plays:

Xizhi Tech: Photonic computing for AI workloads. Replaced electrical interconnects with optical ones for training/inference. Dark pool hit +385%, HK$10k profit per lot. Core thesis: bandwidth density beats copper at scale.

Shengyi Tech: High-layer-count PCB supplier for AI server backplanes. Global #1 in high-performance AI PCB market share. Largest IPO of the year, +50% day one.

Qunhe Tech: Cloud-based 3D design SaaS (China's Coohom equivalent). +144% on debut. Revenue model: freemium → enterprise licensing.

Sunmi IoT: Android-based commercial IoT terminals (those payment/facial recognition screens at retail). Backed by Ant Group, Meituan, Xiaomi. Market pricing it as "next Xizhi" - expects similar multiples.

10 IPOs in April averaged +66% in dark pool trading (HK's pre-listing OTC session, T-1 before official listing).

Access: HK brokerage accounts or USDT-based platforms like StableStock for non-HK residents.

Reality check: This kind of one-way IPO alpha typically evaporates within 2-3 months. Window is closing. Arbitrage opportunities don't stay open when everyone's running the same strat.

Reality check on Cursor Composer after extended use:

Initial impression: Mind-blowing automation potential for coding workflows.

Actual experience after deep usage: Still requires constant human oversight.

Core issues encountered:
- Frequent minor bugs in generated code
- Occasional critical low-level errors that slip through
- Net productivity becomes "three steps forward, one step back"

The problem isn't the tool's capability - it's the accountability gap. When you're responsible for production quality, you can't just ship AI-generated code blindly. You end up context-switching between writing and debugging AI mistakes, which kills the promised efficiency gains.

This is the real bottleneck for AI coding assistants right now: they're powerful co-pilots but unreliable autopilots. The cognitive overhead of quality assurance negates much of the speed advantage.

Lesson learned: Don't bet your entire project architecture on current AI coding tools being production-ready without heavy supervision. The hype cycle hit harder than the actual reliability curve.

Started using Cursor (CC) and initially thought code automation was complete. Reality check after extended use: AI commits frequent small bugs and occasional critical low-level errors.

The workflow bottleneck: You're still responsible for production quality, which means constant code review and cleanup. Net result = three steps forward, one step back.

The takeaway for solo devs: AI coding assistants accelerate prototyping but don't eliminate QA overhead. You're trading typing speed for debugging time. The productivity gain exists but isn't the 10x some claim - more like 2-3x with the quality tax included.

This is why pure AI-generated codebases still need human oversight architecture. The tooling improves iteration speed, but system reliability still requires manual validation layers.

Production infrastructure got nuked by an AI agent that had Railway access. Classic footgun scenario.

Here's the actual defense strategy:

1. Hard isolation: Zero prod credentials on dev machines. All production deployments must route through CI pipelines. If it's on a laptop, it cannot touch prod infrastructure—regardless of what the agent tries to execute.

2. Biometric gating for sensitive operations: If you're early-stage and still need direct prod access, implement pre-execution hooks that require TouchID/FaceID before any infrastructure command runs. A few lines of code can force the agent to pause and request fingerprint auth every time it wants to run terraform, kubectl, or any destructive operation.

The principle: Trust the agent to work, but bioverify before it touches anything critical. Never hand over the keys without a human-in-the-loop gate on destructive actions.

This isn't paranoia—it's basic operational hygiene when your dev tools have autonomous execution capabilities.

OpenAI's 2030 datacenter roadmap targets 30 GW of power capacity. To put that in perspective: 1 GW roughly equals the entire power consumption of Denver, Colorado.

So they're essentially planning to spin up the electrical equivalent of 30 major US cities worth of compute infrastructure in just 5 years.

For context, current hyperscale datacenters typically run 50-150 MW each. A 30 GW target means OpenAI needs 200-600 massive facilities, or a smaller number of truly unprecedented mega-sites.

The grid impact is wild. Most regional power grids supply 5-15 GW total. OpenAI alone would consume more electricity than entire countries like Portugal or Greece.

This isn't just about chips anymore—it's about whether the energy infrastructure can physically scale fast enough to match AI compute ambitions. Nuclear SMRs, dedicated renewable farms, and direct utility partnerships are becoming non-negotiable at this scale.

Here's how to monitor Discord channel messages in Codex and auto-generate daily summaries:

Technical implementation for Discord → Codex integration:

🔧 Core Setup:
• Use Discord's Gateway API or Webhooks to capture real-time channel events
• Stream messages to Codex via their ingestion API
• Set up event listeners for message.create events in target channels

📊 Auto-summarization pipeline:
• Configure scheduled jobs (cron) to aggregate messages by time window
• Feed accumulated messages into an LLM endpoint for summarization
• Output structured daily digests with key discussion points, code snippets, and action items

⚡ Performance considerations:
• Implement rate limiting to avoid Discord API throttling (50 requests per second)
• Use batch processing for message ingestion to reduce API calls
• Cache frequently accessed channel data to minimize latency

💡 Practical use case:
Ideal for dev teams tracking multiple Discord communities, open-source project discussions, or technical support channels without manual monitoring overhead. The summarization layer cuts through noise and surfaces actually relevant technical discussions.

This setup essentially turns Discord into a queryable knowledge base with automated digest generation.