FoundersFeed

Interesting observation on how coding agents might disrupt India's IT outsourcing model. The country built a massive $245B+ software services industry on cost arbitrage - basically being the world's code factory.

But here's the technical shift: AI coding assistants like Cursor, GitHub Copilot, and autonomous agents are commoditizing exactly what made Indian IT shops valuable - writing boilerplate CRUD operations, maintaining legacy systems, and handling repetitive dev work.

The real question is whether Indian firms can pivot fast enough. Some are already integrating AI tooling into their workflows (TCS has its own AI platform, Infosys is pushing generative AI services). But the fundamental economics change when a 10-person team with AI agents can match the output of a 50-person offshore team.

This could actually accelerate India's move up the value chain - forcing companies to focus on complex system architecture, domain expertise, and AI orchestration rather than just code production. The ones that adapt become AI-augmented engineering teams. The ones that don't... well, they're competing with GPT-4 at $0.03 per 1K tokens.

Ex-OpenAI researcher Leopold Aschenbrenner just dropped his Q1 fund holdings. The numbers are absolutely wild:

💰 AUM: $13.6B
📊 Portfolio: 24 positions
📈 YTD Returns: +148%
📉 Max Drawdown: 4.8%

That risk-adjusted return profile (148% gains with sub-5% drawdown) is borderline absurd for a fund this size. For context, most quant funds celebrate 30-40% annual returns with similar drawdown metrics.

What makes this particularly interesting: Aschenbrenner left OpenAI in 2024 after writing controversial internal memos about AGI timelines and security. Now he's running capital with the kind of conviction that suggests he's positioning for the AI infrastructure buildout he predicted.

The 24-position concentration (vs typical 100+ in most tech funds) signals high-conviction thesis-driven allocation rather than diversified beta capture. Would be fascinating to see the actual holdings breakdown - likely heavy on compute infrastructure, AI tooling companies, and semiconductor plays.

Legal update: Musk v. OpenAI case dismissed on statute of limitations grounds.

The court ruled Elon's claims fell outside the allowable timeframe for filing. This means the core arguments about OpenAI's departure from its nonprofit mission and alleged breach of founding agreements never got examined on their merits.

Technically speaking: statute of limitations is a procedural defense that blocks claims filed after a legally defined period expires. In contract disputes, this is typically 2-6 years depending on jurisdiction. The timing of when Musk knew (or should have known) about the alleged breaches was likely the deciding factor.

The dismissal doesn't validate OpenAI's current structure or invalidate Musk's concerns about AGI safety governance. It just means he filed too late. From a legal engineering perspective, this is like trying to merge a PR after the repo has moved on—timing matters as much as the code itself.

Leopold Aschenbrenner (ex-OpenAI researcher) just dropped his SAF fund's Q1 AI portfolio—and it's a massive bet on infrastructure over compute.

Core thesis: Long power + racks + storage. Short logic chips (GPU/ASIC/foundry/lithography).

$8.46B in put options—likely hedging Q1 overheating, not full bearish.

Top 5 holdings:
1. Bloom Energy (BE) — $879M (trimmed)
2. SanDisk (SNDK) — $724M
3. Crown Electrokinetics (CRWV) — $556M (added)
4. Iris Energy (IREN) — $401M (added)
5. Core Scientific (CORZ) — $389M (trimmed)

New positions (3 key directions):
1. Ternium (TE) — domestic solar, power generation angle
2. Shaz — second-tier AI compute play
3. Hive — crypto miner pivoting to AI

Fully exited 3 sub-sectors:
1. Optical comms (Lumentum, Coherent)
2. Natural gas supply chain (EQT, Liberty)
3. Second-tier miners + foundries (Cipher, Hut 8, Tower Semi)

Interesting move: He bought Wu Jihan's BIT Mining.

Translation: He's all-in on data center infrastructure diversification—not betting on single-company execution risk. Power and storage are the new GPUs.

Just shipped GemPod after 7 days of vibe coding - a community platform for Agent Skills discovery and evaluation 🎉

Core features:
• Skill recommendation engine for GitHub open-source agent skills
• Community voting system for quality validation
• Efficient discovery mechanism to filter signal from noise
• Download trend analytics with real user feedback loops
• Real-time update notifications for tracked skills
• Open API for free integration

Tech stack focus: Built for the emerging agent ecosystem where skill composability is becoming critical. Addresses the discoverability problem as LLM agent frameworks proliferate.

Early adopter phase - recruiting founding community members now. This could become the npm/PyPI equivalent for agent capabilities as the space matures.

Polymarket's employee #001 just launched @EntropyMarkets - a compliance-first perpetual futures exchange built on Hyperliquid's HIP-3.

Team breakdown:
- CEO: Ex-Polymarket intern who designed their fee structure and liquidity incentives, then traded at Jump
- Co-founder: Made 8-figure USD doing MEV on Solana, ex-Ribbit Capital
- 8-person team from Jane Street, Jump, Hudson River Trading, Virtu, Radix

Technical approach:
- Starting with Pre-IPO market on HIP-3
- Expanding to equities
- Long-term play: regulated license for US-compliant perpetual swaps

The irony: Hyperliquid just entered prediction markets to compete with Polymarket. Now Polymarket's first hire is building a regulated perps exchange on top of Hyperliquid.

Most on-chain Pre-IPO platforms operate in regulatory gray zones. Entropy's betting that compliance becomes the moat - unsexy but defensible. HIP-3 gives them the infrastructure, regulation gives them the business model differentiation.

Not live yet, but worth tracking given the team's trading infrastructure background.

Polymarket just incubated a stealth team called Entropy Markets that's taking a contrarian bet on regulated perpetual futures.

Team breakdown: 2 Stanford dropouts leading an 8-person crew pulled from Jane Street, Jump, Hudson River Trading, Virtu, and Radix. CEO was Polymarket's first intern who designed their fee structure and liquidity incentives, then traded at Jump. Co-founder pulled 8-figure profits doing MEV on Solana and worked at Ribbit Capital.

Technical play: Building on Hyperliquid's HIP-3 to launch Pre-IPO perpetuals first, then expand to equities. The end goal is getting US regulatory licenses to operate a compliant on-chain perps exchange.

The irony is brutal: Hyperliquid just started moving into prediction markets to compete with Polymarket. Polymarket's response? Spin up a regulated perps exchange that sits ON TOP of Hyperliquid's infrastructure.

Most on-chain Pre-IPO trading operates in regulatory gray zones. Entropy is betting that compliance becomes the actual moat, not the tech itself. Boring? Yes. But regulatory arbitrage is still arbitrage.

This is basically two platforms trying to eat each other's lunch while being forced to dance together on the same chain. Peak crypto game theory.

Hot take: Self-hosting vs managed services in the AI era.

Even though AI can now handle server ops (deployment scripts, monitoring configs, DB tuning), Supabase usage is actually growing. Why?

🔧 Reality check:
- AI writes the code, but YOU still babysit production at 3am
- Postgres replication, backup strategies, and security patches aren't "one prompt away"
- Supabase abstracts: realtime subscriptions, auth flows, edge functions, storage APIs

💰 Economics:
- Your time debugging self-hosted infra > $25/month managed tier
- AI reduces setup friction, but ongoing ops cost remains

🎯 The pattern:
AI lowers the barrier to self-host for learning/side projects, but production workloads still favor managed services for reliability and velocity.

TLDR: AI makes you dangerous enough to deploy, not experienced enough to run it 24/7.

Interesting observation: AI can now handle server ops pretty well, so why is Supabase usage actually growing?

Here's the technical reality:

1. AI ops still require expertise to set up properly - you need to configure monitoring, define runbooks, handle edge cases. Supabase abstracts all of that instantly.

2. Time-to-production matters more than ever. Spinning up Postgres + Auth + Storage + Realtime with AI takes hours of prompting and debugging. Supabase does it in 2 minutes.

3. AI can manage infrastructure, but it can't guarantee SLA, handle DDoS, or provide compliance certifications out of the box.

4. The real bottleneck shifted: developers now spend time on product logic, not server config. AI coding assistants amplify this - you want to ship features, not babysit infrastructure.

5. Cost efficiency at scale: Supabase's shared infrastructure model is often cheaper than running dedicated instances, even with AI optimization.

The paradox is that AI made developers MORE productive at writing code, which makes managed services MORE valuable because they remove the last remaining friction. You're not paying for something AI could do - you're paying to not think about it at all.

Anthropic just pulled off something wild: they rewrote Bun in Rust in 6 days using AI agents. 6,755 commits, zero human-written code, zero human code review, force-merged straight to production.

This is a massive shift in how software gets built. The entire codebase transition happened autonomously - no developer manually typing Rust, no PR reviews, just AI agents churning through the rewrite and pushing it live.

What makes this significant:
- Speed: 6 days for a complete language migration that would normally take months
- Scale: Nearly 7k commits without human intervention
- Trust model: Running code in prod that humans haven't manually verified

This raises serious questions about code ownership, debugging, and maintenance. When bugs surface, who debugs AI-generated Rust? How do you reason about performance regressions in code no human wrote?

The era of "if you can't explain it, don't ship it" might be ending. We're entering a phase where AI-generated codebases run critical infrastructure, and humans become operators rather than authors.

CXMT (ChangXin Memory Technologies) just dropped their IPO filing with insane Q1 2026 numbers: ¥50.8B revenue (up 719% YoY), ¥33B net profit — that's 20x their entire 2025 annual profit. Operating margin hit 65%, trailing only SK Hynix (72%) and Micron (75%+).

The DRAM supercycle is here. AI infrastructure demand for HBM is so intense that Samsung/SK Hynix/Micron are redirecting capacity from commodity DRAM (DDR4/DDR5) to HBM production. Result: DDR contract prices spiked 58-63% QoQ in Q1 2026. CXMT, sitting at 4.7% global market share (#4 worldwide), is cashing in on the commodity shortage while the Big 3 chase HBM margins.

But here's the brutal reality: CXMT has zero HBM production. They're mass-producing DDR4/DDR5/LPDDR4/5 on 16nm process (no EUV), while SK Hynix and Samsung are already shipping HBM4. CXMT's HBM3 won't hit volume until late 2026 at their Shanghai packaging facility — that's a 2-generation gap. Their 16nm vs competitors' 12nm represents a 3-4 year process node deficit without EUV lithography access.

TrendForce projects HBM will grow from 12% of DRAM revenue (2024) to 35%+ by 2027, with total DRAM market approaching $400B. Translation: if you're not in HBM by 2027, you're stuck in the low-margin commodity ghetto.

CXMT's window is tight. If they can ramp HBM3 production and fast-track HBM4 development, they might crack the "HBM club" by 2030 and legitimately challenge the oligopoly. Miss that window, and they're permanently relegated to niche commodity DRAM supplier status — profitable today, irrelevant tomorrow. The next 18 months will determine if China gets a seat at the high-performance memory table or watches from the sidelines.

GemPod just launched - a rating/review platform specifically for Agent Skills. 🎉

The problem: Agent Skills protocol has been open for 5 months. Teams solved installation issues, some tackled security reviews, but nobody addressed quality evaluation of these skills.

GemPod's approach:
- User recommendations + human voting system to surface actually useful skills
- Categorized skill directory with GitHub stars/forks filtering
- Built-in EN/CN translation
- Stars history charts integration
- Skill following & badge system

Currently in Beta with gated registration:
- Auto-approved: GitHub repo with 500+ stars OR X account with 1000+ followers
- Otherwise: apply on-site or DM for invite code

The gap this fills: With thousands of Agent Skills floating around, there's no signal vs noise filter. GemPod aims to be that curation layer through community validation rather than just star counts.

Seedance 2.0 drops with single-prompt video editing that actually works.

Core capability: Object replacement, environment swaps, and seasonal transformations without re-rendering from scratch. Feed it your existing AI video + one text prompt, and it handles the rest.

Technical angle: This is likely using temporal consistency models combined with inpainting techniques—maintaining motion coherence while swapping visual elements. The fact it preserves the original video's motion data while manipulating specific regions suggests some form of latent space manipulation rather than full regeneration.

What makes this interesting: Most AI video tools force you to regenerate entire sequences when you want changes. Seedance 2.0 appears to do localized edits while keeping temporal flow intact. That's the hard part—maintaining frame-to-frame consistency during edits.

Practical use cases: Rapid A/B testing for video ads, seasonal content variations without reshoots, quick object replacements for concept validation.

The real test: How well it handles complex motion, occlusions, and lighting changes. Those are where most video editing models fall apart.

Seedance 2.0 just dropped character swapping and it's genuinely impressive - single generation, no iteration hell.

The tech eliminates the traditional multi-pass workflow. Instead of generating base → extracting features → re-injecting → hoping for consistency, Seedance 2.0 does character replacement in one shot.

Why this matters: Most character-consistent generation tools require 3-5 iterations minimum. This collapses that entire pipeline into a single inference pass, which means faster prototyping and lower compute costs.

Likely using some form of spatial conditioning with identity preservation at the latent level - similar to IP-Adapter but with tighter control over character features during the diffusion process.

For anyone building character-driven content (games, animation pipelines, storytelling tools), this is a serious workflow upgrade. No more prompt engineering gymnastics to maintain character consistency across scenes.

Bill Gates has completely exited his Microsoft position. Gates Foundation's Q1 SEC filing shows they dumped their final 7.7M MSFT shares (~$3.2B). Zero stake remaining in the company he founded.

Technically significant because:
• This marks the end of founder ownership in one of tech's most valuable companies (current market cap ~$3T)
• Gates had been gradually reducing his position since stepping down from the board in 2020
• The foundation's portfolio strategy now appears fully decoupled from Microsoft's trajectory
• Timing is interesting - comes as MSFT is heavily investing in OpenAI infrastructure and Azure AI buildout

For context: Gates left Microsoft's board in 2020 to focus on philanthropy. His stake had already shrunk from ~24% at IPO (1986) to ~1.3% by 2019. This final exit completes a multi-decade divestment process, likely driven by foundation diversification requirements rather than any technical concerns about MSFT's AI positioning.

GPT-4.5 medium is now handling most dev tasks surprisingly well. The standout feature? Output speed is significantly faster than previous models.

Question for the community: What reasoning level are you running GPT-4.5 at for your workflows? Curious to see if higher reasoning tiers are worth the latency trade-off or if medium hits the sweet spot for most coding tasks.

GPT-5.5 Medium is now handling most dev tasks surprisingly well. The real standout? Output speed is blazing fast. OpenAI continues to deliver on performance where it counts for actual development workflows.

ChatGPT's financial integration just went from read-only to write-capable. The implications are wild:

Current state: passive data reading (account balances, transaction history)

Next phase: autonomous financial operations
- Auto-rebalancing portfolios into lower-expense-ratio index funds
- Balance transfer optimization across credit cards to maximize 0% APR periods
- CD laddering with real-time rate arbitrage across institutions

This isn't a fintech play, it's a full-stack assault on consumer finance. OpenAI is positioning to disintermediate:
- Robo-advisors (Betterment, Wealthfront)
- Personal finance apps (Mint, YNAB)
- Traditional banking UX layers

Second-order effect that's getting zero attention: advertising revenue collapse. Financial services companies dump $15B+ annually into customer acquisition. If an AI agent handles fund allocation decisions autonomously, those ad budgets evaporate overnight. No more credit card comparison shopping, no more "best savings account" searches.

The endgame: ChatGPT becomes the default financial decision layer for consumers, and traditional finance companies get relegated to dumb infrastructure providers competing purely on rates and fees. Distribution power shifts entirely to the AI layer.

Interesting geopolitical tech question: Would you trade access to latest Nvidia chips (H100/H200 class) to China in exchange for mutual AI safety verification?

The proposal: Reciprocal lab monitoring - Chinese safety observers embedded in US AI labs, US observers in theirs. Essentially treating frontier AI development like nuclear inspection protocols.

Technical implications:
- China currently limited to A800/H800 (crippled versions with reduced NVLink bandwidth)
- Full H100 access would dramatically accelerate their LLM training capabilities
- Question is whether bilateral safety oversight creates enough risk mitigation

The core tension: Hardware export controls are currently the main lever slowing China's frontier model development. Trading that leverage for inspection rights assumes:
1. Safety protocols can be meaningfully verified
2. Both sides operate in good faith
3. The inspection regime can detect violations before they matter

This essentially asks: Is asymmetric compute advantage more valuable than symmetric safety transparency?

What's your take - is this a reasonable framework or naive wishful thinking about AI governance?

ChatGPT mobile now supports Codex integration 📱

You can now run code execution and debugging directly from the ChatGPT mobile app. This means you can test Python scripts, analyze data, and execute code snippets without needing a laptop.

Technically significant because:
- Brings full Code Interpreter capabilities to mobile
- Enables on-the-go development and debugging
- Same sandboxed Python environment as desktop
- Can handle file uploads and generate downloadable outputs

Practical use cases: quick data analysis, testing algorithms, debugging code snippets, running automation scripts while mobile. The execution environment is identical to desktop, so you get the same timeout limits and package availability.

This effectively turns your phone into a portable Python REPL with GPT assistance.