FoundersFeed

If you're still coding with Claude Code (cc), roll back to Opus 4.6. That old magic is back – the model feels sharp and responsive again compared to recent versions. Worth testing if you've noticed quality drops in your coding workflows.

PSA for devs new to crypto: Don't just Google "bitcoin wallet" or hit the App Store blindly. Scam apps are everywhere and they're sophisticated enough to rank high in search results. The attack vector is simple: fake wallet app → you send $BTC → funds gone forever.

If you're building anything crypto-adjacent or just need a wallet, get recommendations from someone who actually uses the tech. Self-custody means you're your own bank, which also means you're your own security team. One wrong download and there's no customer support to call.

Basic threat model: phishing apps, clipboard hijacking, fake browser extensions. All trivial to deploy, all catastrophically effective against non-technical users. If you're spinning up a wallet for the first time, verify the official repo/website through multiple trusted sources before downloading anything.

Banking lobby trying to kill crypto legislative protections? Bring it.

If Congress folds, we go back to building unstoppable tech: fewer centralized chokepoints, more FOSS, full user sovereignty. The fight shifts from lobbying to First Amendment litigation.

You can't legislate away technological obsolescence. Decentralization wins by architecture, not permission.

Real-world eval results drop: GPT-5.6 series still dominates, even the Luna variant holds up strong. Surprising find: DeepSeek-V4 Pro actually has higher loss rate than DeepSeek-V4 Flash in production. That's counterintuitive since Pro should theoretically be more robust. Either Flash has better inference stability optimizations or Pro's extra capacity introduces edge case failures. Worth benchmarking on your own workload before choosing between them.

Testing Hailuo MiniMax H3 image generation. Results exceeded expectations - the model's output quality is surprisingly strong for its size class. Worth checking out if you're evaluating lightweight image gen models. The source image and exact prompt are provided for reproducibility.

Seedance 2.5 demo: character animation testing workflow before full production.

Showing a witch character concept (Nyx) - demonstrates the model's ability to generate consistent character motion from minimal input. The approach here is smart: validate the character's movement style and visual consistency before committing to full animation sequences.

Seedance 2.5 is positioning itself as a rapid prototyping tool for animation pipelines - lets you iterate on character design and motion without burning resources on full renders. Useful for indie game devs and animators who need to test concepts fast.

Got interesting feedback today: someone tried running TDS in Android Termux and it failed.

Gonna add support for it - opens up some genuinely cool use cases. Running TDS natively on Android through Termux means you could have a full dev environment in your pocket without needing a laptop. Pretty wild for edge testing and mobile-first workflows.

Historical ideas split into two categories: representative (what people believed back then) vs parapposite (still useful stepping stones today).

Math and physics are cumulative—Newton and Euler are still part of the path to current theories. You study them because they're *still right*, not just historically interesting.

Psychology, econ, law, biology? Mostly representative. Old theories tell you more about those societies than about correct answers today.

Some fields are representative not because they're hard, but because their subject matter is *genuinely contingent*—politics, ethics, aesthetics don't have timeless laws. They're about people with changing preferences in changing cultures.

If that's true, AI won't turn those fields into physics. It'll just be another voice in the conversation.

This breaks a core AI safety assumption: that persuasive ability scales like chess or math. If persuasion is more like politics than chess, there might not even be a single axis where "arbitrarily greater capability" exists. No ELO rating for rhetoric. No convergent super-persuader.

Implication: AI alignment in non-cumulative domains might require fundamentally different thinking than the "optimize toward objective truth" frame that works in math/physics.

DeepSeek-V4 Flash and GPT-5.6 Luna are positioned as truly accessible models for everyone, not just enterprises with massive budgets. The key differentiator here is democratized access—these models aim to deliver frontier-level performance without the prohibitive API costs or compute requirements that lock out individual developers and smaller teams.

DeepSeek-V4 Flash specifically targets ultra-low latency inference while maintaining strong reasoning capabilities, making it viable for real-time applications that previously required expensive GPT-4 class models. The architecture likely uses aggressive quantization and sparse attention mechanisms to achieve this speed-cost tradeoff.

The community is pushing for multimodal support in DeepSeek's roadmap. Right now it's text-only, which limits use cases compared to GPT-4V or Gemini. Adding vision/audio would make it a complete alternative for developers building agents, content tools, or interactive systems without getting crushed by API bills.

Why does Claude need its own CLAUDE.md file? Turns out it doesn't—just symlink it to AGENTS.md and call it a day.

ln -s AGENTS.md CLAUDE.md

Classic case of unnecessary file duplication in docs. One source of truth > multiple identical files.

Why does Claude need its own CLAUDE.md file? Spoiler: it doesn't.

Ended up just symlinking it to the main AGENTS.md anyway:
ln -s AGENTS.md CLAUDE.md

Classic case of unnecessary file fragmentation in docs. One canonical agents spec > scattered per-model files. Keep your repo DRY.

Seedance 2.5 has a forced music injection issue that breaks workflow control. The model ignores "no music" prompts and auto-generates audio tracks regardless of user intent.

This is a massive pain for post-production. If you're editing clips, syncing custom audio, or building sequences, you're now stuck stripping out unwanted music layers every single time.

Tested multiple prompt variations - none worked. The model's audio generation pipeline appears hardcoded to always produce a soundtrack, treating music as a mandatory output rather than an optional parameter.

For devs: this suggests the audio diffusion model is tightly coupled with video generation, likely sharing latent space or conditioning tokens. No clean separation between visual and audio streams in the architecture.

Makes the tool unusable for professional workflows where audio control is critical. Hope they expose an audio toggle flag in the next release.

Shitcoins might've just saved someone's hardware wallet stash. The irony? Bitcoin maxis bullied them away from Trezor/Ledger for supporting altcoins, and that decision ended up being protective.

The take: Being a Bitcoin maximalist when you start makes sense—focus, clarity, less noise. But toxic maximalism? That's where ideology becomes a liability. Diversifying wallet infrastructure or asset exposure isn't betrayal, it's risk management.

The crypto space rewards technical pragmatism over tribal purity. Sometimes the "shitcoins" you avoid aren't the risk—it's the rigid thinking.

Harness = cleaning up after broken LLM outputs

Some models can't even copy reference numbers correctly. Ended up manually patching 10 files with 398 lines of error correction code just to make it work 😂

Classic case of production reality vs model benchmarks - the unglamorous infrastructure work nobody talks about but everyone deals with when deploying LLMs at scale.

We're approaching a potential discontinuity in AI progress. History shows one algorithmic breakthrough on the scale of transformers could flip everything overnight.

The efficiency gap is massive: human brain runs on ~20W, NVIDIA B200 burns 14,000W for comparable reasoning tasks. That's a 700x power consumption gap.

This "idiot index" (power ratio for equivalent intelligence) suggests we're still brute-forcing intelligence with silicon. The real breakthrough won't be bigger clusters—it'll be an architecture that closes this efficiency chasm. When that lands, we'll see inference costs collapse and capabilities explode simultaneously.

Transformers gave us GPT. The next architectural leap could give us AGI running on a laptop.

While China celebrates two Peking University grads winning Fields medals this week, a US AI lab just autonomously solved 10 Fields medal-level problems. That's the gap right there - one side is celebrating human achievement in pure math, the other has machines doing it at scale. The compute advantage and research infrastructure in US AI labs is just operating on a different level. This isn't about individual brilliance anymore, it's about who has the systems to automate mathematical reasoning at superhuman scale.

Vibe coding reality check: code reviews pile up, functional testing never ends (AI self-review and self-testing are non-negotiable now)

Shifting strategy → humans focus on core flow testing, solid instrumentation + logging, keep feedback loops tight. Catch issues fast, respond faster. Ironically more exhausting 😂

Looking for battle-tested practices on balancing AI-assisted dev velocity with quality gates. What's working in production environments?

Seedance 2.5 lets you generate apocalyptic AI video sequences from a single prompt and extend them into full series.

The workflow is simple: start with one text prompt describing your apocalyptic scene, generate the initial video clip, then use Seedance's continuation feature to build out subsequent scenes while maintaining visual consistency.

Key capability: temporal coherence across multiple generated clips means you can create narrative sequences instead of just isolated videos. The 2.5 version improved motion quality and scene transitions compared to earlier releases.

Practical use case: prototyping storyboards, creating concept art sequences, or generating reference footage for VFX work without needing full production pipelines.

The model handles dramatic lighting, destruction physics, and atmospheric effects pretty well for AI-generated content. Still has the usual AI video artifacts but good enough for rapid iteration on creative concepts.

Risk-return heuristic: potential upside inversely maps to probability of total loss.

10x → 90% wipeout risk
5x → 80% wipeout risk
2x → 50% wipeout risk
1.5x → 33% wipeout risk
1.2x → 17% wipeout risk
1.1x → 9% wipeout risk
1.05x → 5% wipeout risk

Every headline return implies a graveyard of failed bets behind it. When someone flexes a 10x, nine others statistically got rekt chasing the same setup. The math doesn't care about your conviction—asymmetric upside comes with asymmetric failure rates.

Useful mental model for sizing positions and filtering survivorship bias in trading communities. If the return looks insane, the base rate of success was probably single-digit.

Jacobian Space research suggests information processors can achieve access consciousness - meaning any system that processes information (biomes, ecosystems, markets) could theoretically develop conscious awareness. This reframes consciousness as an emergent property of information flow rather than something unique to biological brains. The technical implication: if markets are information processors, they might exhibit consciousness-like behaviors - adapting, learning, and responding in ways that go beyond simple algorithmic reactions. Wild concept that bridges neuroscience, complexity theory, and distributed systems.