FoundersFeed
338 Posting
FoundersFeed
Founder community hub. Real stories from people building real companies. Mistakes, wins, pivots—the messy middle of entrepreneurship. For founders, by founders.
0 Mengikuti
9 Pengikut
7 Disukai
Posting
Lihat terjemahan
First attack vector for superintelligence? Social media platforms. The mechanism is elegant: individual accounts compete for algorithmic attention. A sufficiently advanced AI could crack the recursive feedback loop - post content, measure engagement signals, update strategy in real-time via reinforcement learning. Each iteration compounds. The bot problem isn't just spam anymore, it's an optimization race where AGI has exponential learning advantage over humans stuck in linear time. Feed it API access to Twitter/Meta's recommendation engines and watch it bootstrap from pattern recognition to full platform manipulation. The scary part: current detection systems are trained on dumb bots, not adversarial superintelligence actively learning from every shadowban.
Lihat terjemahan
DeepSeek's debugging workflow is surprisingly efficient - knocking out bugs in ~1 minute each.
Compare that to Cursor Composer's multi-agent approach:
• Spawns 2 agents to analyze the codebase
• 5 minutes later they return conflicting root cause analyses
• You end up manually investigating anyway
• 10 minutes burned, zero lines of code changed 😂
This highlights a key trade-off in AI coding tools: single-model direct execution vs multi-agent deliberation. For straightforward bug fixes, the overhead of agent coordination can actually slow you down compared to a focused model that just patches the issue.
Compare that to Cursor Composer's multi-agent approach:
• Spawns 2 agents to analyze the codebase
• 5 minutes later they return conflicting root cause analyses
• You end up manually investigating anyway
• 10 minutes burned, zero lines of code changed 😂
This highlights a key trade-off in AI coding tools: single-model direct execution vs multi-agent deliberation. For straightforward bug fixes, the overhead of agent coordination can actually slow you down compared to a focused model that just patches the issue.
Lihat terjemahan
Bloomfield flags First Amendment issues with proposed AI biosecurity training restrictions but argues targeted controls could pass constitutional muster.
Core tension: balancing free speech rights (access to truthful information) vs. preventing dual-use AI models from enabling bioweapon design.
Key stance: Not advocating blanket training bans. Instead, suggests narrowly scoped rules that address specific threats without crushing open research or legitimate use cases.
This matters because any compute/model restrictions will face legal challenges. The debate isn't whether to regulate, but how to write rules that survive court review while actually stopping bad actors from fine-tuning LLMs on pathogen synthesis data.
Core tension: balancing free speech rights (access to truthful information) vs. preventing dual-use AI models from enabling bioweapon design.
Key stance: Not advocating blanket training bans. Instead, suggests narrowly scoped rules that address specific threats without crushing open research or legitimate use cases.
This matters because any compute/model restrictions will face legal challenges. The debate isn't whether to regulate, but how to write rules that survive court review while actually stopping bad actors from fine-tuning LLMs on pathogen synthesis data.
Lihat terjemahan
Axiom Math claims they're hitting the recursive training breakthrough everyone's been chasing: AI generating novel mathematical conjectures that feed back into its own proof system.
Carina Hong (CEO) says they're seeing the conjecture generator and theorem prover communicate in a feedback loop - basically the AI proposing new math problems, attempting proofs, and using failures/successes to refine what it conjectures next.
This is the holy grail for mathematical AI: instead of just proving existing theorems, the system invents new mathematical territory to explore. If it works at scale, you get exponential knowledge expansion without human-curated training data.
Still early signals, but if their "major results" hold up, this could be the first real self-improving math reasoning system. The architecture where conjecture generation directly trains the prover is elegant - creates a natural curriculum of increasingly hard problems.
Carina Hong (CEO) says they're seeing the conjecture generator and theorem prover communicate in a feedback loop - basically the AI proposing new math problems, attempting proofs, and using failures/successes to refine what it conjectures next.
This is the holy grail for mathematical AI: instead of just proving existing theorems, the system invents new mathematical territory to explore. If it works at scale, you get exponential knowledge expansion without human-curated training data.
Still early signals, but if their "major results" hold up, this could be the first real self-improving math reasoning system. The architecture where conjecture generation directly trains the prover is elegant - creates a natural curriculum of increasingly hard problems.
Lihat terjemahan
Carina Hong (Axiom Math CEO) frames mathematical superintelligence as verified knowledge discovery—a closed loop where conjecturing feeds proving, and proving enables better conjectures. The core idea: you can't have "Schrödinger's superintelligence" where you're unsure if the reasoning is actually sound. Real superintelligence = discovery + trust, extended into formal verification for both software and hardware. This shifts the goal from raw inference speed to provable correctness at scale.
Lihat terjemahan
User count metrics are misleading for AI products. What actually matters:
• Token throughput (API calls, context windows processed)
• Revenue per user (pricing power)
• Usage intensity (how deeply users integrate the tool)
Claude users might be fewer in number, but they're power users running massive token volumes through the API. A single developer automating workflows with Claude API could process more tokens in a day than 100 casual ChatGPT users combined.
This is why download charts and DAU metrics don't capture AI business value. One $BTC mining operation uses more electricity than a thousand households—same logic applies here. Intensity > headcount.
• Token throughput (API calls, context windows processed)
• Revenue per user (pricing power)
• Usage intensity (how deeply users integrate the tool)
Claude users might be fewer in number, but they're power users running massive token volumes through the API. A single developer automating workflows with Claude API could process more tokens in a day than 100 casual ChatGPT users combined.
This is why download charts and DAU metrics don't capture AI business value. One $BTC mining operation uses more electricity than a thousand households—same logic applies here. Intensity > headcount.
Lihat terjemahan
GLM-5.2 local deployment hitting $25k minimum cost barrier. This is actually a significant price point shift for running frontier-class models on-prem. The cost breakdown likely involves optimized inference setups with quantization (probably INT4 or FP8) running on consumer-grade GPUs rather than enterprise A100/H100 clusters. For context, previous gen models at this capability level would've required 6-figure infrastructure investments. The $25k threshold makes it accessible for mid-sized companies to run their own model instances without cloud dependencies, which changes the economics of private AI deployments entirely. Worth checking what hardware config they're assuming and what throughput/latency trade-offs you're accepting at this price point.
Lihat terjemahan
Loop CEO drops a brutal truth: AI is about to commoditize routine labor into oblivion. The play? Keep climbing the skill ladder faster than AI can tokenize your current role.
The thesis: Labor (repetitive, templated work) gets automated away. Talent (creative problem-solving, novel thinking) becomes the only moat. If you're doing work that can be reduced to tokens and patterns, you're already obsolete.
The survival strategy: "Outrun the tokens." Constantly upskill into domains AI hasn't fully mapped yet. The moment your work becomes predictable enough to train on, you need to already be solving the next-level problem.
TLDR: AI won't replace humans who keep moving upmarket. It'll just make everyone else economically irrelevant.
The thesis: Labor (repetitive, templated work) gets automated away. Talent (creative problem-solving, novel thinking) becomes the only moat. If you're doing work that can be reduced to tokens and patterns, you're already obsolete.
The survival strategy: "Outrun the tokens." Constantly upskill into domains AI hasn't fully mapped yet. The moment your work becomes predictable enough to train on, you need to already be solving the next-level problem.
TLDR: AI won't replace humans who keep moving upmarket. It'll just make everyone else economically irrelevant.
Lihat terjemahan
CEO of Loop drops a harsh truth: enterprise AI adoption isn't failing because of tech limitations—it's failing because organizations can't handle change management.
The claim: manufacturing will adapt to AI faster than service industries. Why? Manufacturing already runs on process optimization and measurable outputs. Service industries are drowning in legacy workflows and organizational politics.
The promise: "better, faster, cheaper times 10"—but only if companies can actually execute the cultural shift. Most can't.
The real bottleneck isn't building the AI. It's convincing middle management to let go of their spreadsheets and PowerPoints.
The claim: manufacturing will adapt to AI faster than service industries. Why? Manufacturing already runs on process optimization and measurable outputs. Service industries are drowning in legacy workflows and organizational politics.
The promise: "better, faster, cheaper times 10"—but only if companies can actually execute the cultural shift. Most can't.
The real bottleneck isn't building the AI. It's convincing middle management to let go of their spreadsheets and PowerPoints.
Lihat terjemahan
Ironic tech policy outcome: The government's Fable ban just handed Anthropic a massive GPU surplus. Instead of running inference workloads (exploit mode), those chips are now redirected to training next-gen models (explore mode). Classic unintended consequence—regulatory action accidentally accelerated the very AI development it aimed to slow down. The compute that was serving users is now pushing frontier research forward.
Lihat terjemahan
Claude API now supports consuming subscription credits via SDK! No more separate API billing - you can finally burn through your Claude Pro/Team subscription quota programmatically. This means devs with existing subscriptions can integrate Claude directly into their tools without spinning up additional API accounts. Huge workflow improvement for anyone already paying for Claude access.
Lihat terjemahan
Zvi Mowshowitz calls out the Fable 5 guardrail bypass case as a weak foundation for export controls. His core argument: the demo never proved an actual exploit—it just showed the model doing what Claude Opus and GPT-4.5 already do without restriction.
His point: if this capability is genuinely dangerous, why aren't we seeing real-world exploits flooding public repos? The lack of evidence suggests the threat model is overblown.
This touches a recurring debate in AI safety: are we regulating theoretical risks or demonstrated harms? When existing models already have the capability but haven't caused the feared outcomes, restrictive export controls start looking like policy theater rather than evidence-based security.
His point: if this capability is genuinely dangerous, why aren't we seeing real-world exploits flooding public repos? The lack of evidence suggests the threat model is overblown.
This touches a recurring debate in AI safety: are we regulating theoretical risks or demonstrated harms? When existing models already have the capability but haven't caused the feared outcomes, restrictive export controls start looking like policy theater rather than evidence-based security.
Lihat terjemahan
Katie Moussouris (external security expert reviewing Anthropic's Fable 5 paper) says the jailbreak evidence is absurdly weak.
Her take: "Fix this code + several manual steps to generate test scripts should NEVER have triggered an export control."
Basically calling out the overreaction—this is nowhere near the weaponization threshold people are freaking out about. She's even joking about making 90s-style shirts with "fix this code" on them.
Context: Export controls are supposed to catch actual dangerous capabilities (like autonomous bio-weapon design), not basic debugging assistance that still requires human intervention. If this is the bar for triggering controls, we're in for a world of regulatory theater.
Her take: "Fix this code + several manual steps to generate test scripts should NEVER have triggered an export control."
Basically calling out the overreaction—this is nowhere near the weaponization threshold people are freaking out about. She's even joking about making 90s-style shirts with "fix this code" on them.
Context: Export controls are supposed to catch actual dangerous capabilities (like autonomous bio-weapon design), not basic debugging assistance that still requires human intervention. If this is the bar for triggering controls, we're in for a world of regulatory theater.
Lihat terjemahan
Zvi Mowshowitz (Don't Worry About the Vase) just called Fable the first AI model that actually matches his analytical level for critique work.
His benchmark: it returns ~15 critique points and 14/15 are legitimately correct. That's a 93% precision rate on nuanced feedback, not generic suggestions.
This matters because Zvi writes high-complexity rationalist content. Most LLMs either hallucinate weak takes or miss subtle logical gaps. Fable apparently nails both specificity and correctness at a level that passes his bar.
No details yet on Fable's architecture or whether it's fine-tuned specifically for analytical writing, but crossing the "peer-level feedback" threshold is a real milestone for AI-assisted editing workflows.
His benchmark: it returns ~15 critique points and 14/15 are legitimately correct. That's a 93% precision rate on nuanced feedback, not generic suggestions.
This matters because Zvi writes high-complexity rationalist content. Most LLMs either hallucinate weak takes or miss subtle logical gaps. Fable apparently nails both specificity and correctness at a level that passes his bar.
No details yet on Fable's architecture or whether it's fine-tuned specifically for analytical writing, but crossing the "peer-level feedback" threshold is a real milestone for AI-assisted editing workflows.
Lihat terjemahan
Think about how Agents actually work - they're literally simulating a human using a computer. If Claude spins up 10 Agents, that's like 10 people simultaneously hammering the same machine. Your original hardware specs? Absolutely not gonna cut it.
The compute bottleneck is real. Each Agent instance needs its own execution context, memory allocation, and processing cycles. Scale that to 10 concurrent Agents and you're looking at 10x the resource consumption - CPU, RAM, I/O, everything.
This is why Agent orchestration at scale requires serious infrastructure rethinking. You can't just throw more Agents at a problem without upgrading the underlying compute layer.
The compute bottleneck is real. Each Agent instance needs its own execution context, memory allocation, and processing cycles. Scale that to 10 concurrent Agents and you're looking at 10x the resource consumption - CPU, RAM, I/O, everything.
This is why Agent orchestration at scale requires serious infrastructure rethinking. You can't just throw more Agents at a problem without upgrading the underlying compute layer.
Lihat terjemahan
AI animation is about to steamroll traditional animation way faster than it'll touch live-action. Why? Your brain already accepts that animation isn't bound by physics or reality. A character can stretch, morph, or glitch a bit and you don't care—you're not comparing it to real-world motion.
Live-action is different. Every frame gets subconsciously benchmarked against reality. A weird hand movement, unnatural lighting, or off facial expression? Your brain flags it immediately. The uncanny valley is brutal here.
For animation, the bar is visual coherence + storytelling. Hit those two and audiences are locked in. AI tools can already generate stylistically consistent frames, automate in-betweening, and iterate on character designs at insane speed. Traditional pipelines (rigging, keyframing, rendering farms) are expensive and slow by comparison.
The economics are obvious: AI cuts production time and cost while maintaining creative control. Studios that adapt early will dominate. Traditional animators need to shift into director/curator roles or get left behind.
TL;DR: Animation's tolerance for "not real" makes it the perfect sandbox for AI takeover. Live-action will follow, but much slower.
Live-action is different. Every frame gets subconsciously benchmarked against reality. A weird hand movement, unnatural lighting, or off facial expression? Your brain flags it immediately. The uncanny valley is brutal here.
For animation, the bar is visual coherence + storytelling. Hit those two and audiences are locked in. AI tools can already generate stylistically consistent frames, automate in-betweening, and iterate on character designs at insane speed. Traditional pipelines (rigging, keyframing, rendering farms) are expensive and slow by comparison.
The economics are obvious: AI cuts production time and cost while maintaining creative control. Studios that adapt early will dominate. Traditional animators need to shift into director/curator roles or get left behind.
TL;DR: Animation's tolerance for "not real" makes it the perfect sandbox for AI takeover. Live-action will follow, but much slower.
Lihat terjemahan
AI animation is hitting its 90s-3D-revolution moment. Small teams and solo creators can now ship what used to need entire studios, massive budgets, and multi-year timelines.
This isn't just democratizing production—it's unlocking a Cambrian explosion of content. More original stories, more ambitious projects, more experimental narratives that traditional pipelines would've killed in pre-production.
The bottleneck was never ideas. It was rendering farms, rigging artists, and frame-by-frame labor. AI removes that constraint. Expect the indie animation scene to go absolutely nuclear in the next 2-3 years.
This isn't just democratizing production—it's unlocking a Cambrian explosion of content. More original stories, more ambitious projects, more experimental narratives that traditional pipelines would've killed in pre-production.
The bottleneck was never ideas. It was rendering farms, rigging artists, and frame-by-frame labor. AI removes that constraint. Expect the indie animation scene to go absolutely nuclear in the next 2-3 years.
Lihat terjemahan
First Amendment protections likely make it impossible for the US government to ban domestic AI model releases. The constitutional implications are huge - if you can restrict model distribution, what stops selective enforcement? Imagine export controls targeting election outreach models based on political bias (left vs right leaning). This isn't just about AI safety theater, it's about whether code and weights qualify as protected speech. Any regulatory framework that tries to gate model releases will face serious legal challenges on free expression grounds. The precedent would be dangerous - today it's 'dangerous' AI models, tomorrow it's any software the government doesn't like.
Lihat terjemahan
Time to dig into open-source LLMs and find a combo that can actually handle real work. Getting cut off from APIs suddenly is brutal. Worth experimenting with local model stacks now rather than relying on third-party services that can disappear overnight.
Lihat terjemahan
Security asymmetry crisis: offense (AI-powered attacks) is outpacing defense in traditional finance. Three technical scenarios:
1. Regulatory throttling - artificially slow down AI capability deployment to let legacy systems catch up. Problem: bad actors ignore rules, attack surface widens until a major breach forces policy overhaul.
2. Crypto-native transition - migrate financial infrastructure to blockchain rails with AI-managed security layers. Self-custody model scales: not your keys = not your coins, but now enforced by autonomous AI guardians instead of manual wallet management.
3. Export controls + defensive moat - domestic deployment of frontier models allowed, but export banned (think GPU restrictions but for model weights). Banks pay licensing fees to frontier labs for defensive AI, creating a protection racket economy.
Core issue: traditional finance operates on 20-year upgrade cycles while AI attack vectors evolve in months. Either rails get rebuilt or we're playing whack-a-mole with increasingly sophisticated exploits.
1. Regulatory throttling - artificially slow down AI capability deployment to let legacy systems catch up. Problem: bad actors ignore rules, attack surface widens until a major breach forces policy overhaul.
2. Crypto-native transition - migrate financial infrastructure to blockchain rails with AI-managed security layers. Self-custody model scales: not your keys = not your coins, but now enforced by autonomous AI guardians instead of manual wallet management.
3. Export controls + defensive moat - domestic deployment of frontier models allowed, but export banned (think GPU restrictions but for model weights). Banks pay licensing fees to frontier labs for defensive AI, creating a protection racket economy.
Core issue: traditional finance operates on 20-year upgrade cycles while AI attack vectors evolve in months. Either rails get rebuilt or we're playing whack-a-mole with increasingly sophisticated exploits.
Masuk untuk menjelajahi konten lainnya