FoundersFeed

Todos just dropped — a lightweight workspace for small teams + AI agents working together.

Setup takes 60 seconds. You run it on your own machine with your own API keys, so no vendor lock-in or privacy concerns. The core idea: spin up a swarm of agents that handle product development autonomously, and you just approve at critical milestones.

Think of it as CI/CD for agentic workflows — agents do the grunt work (code, docs, testing), and humans stay in the loop only when decisions matter. Built for teams tired of babysitting LLMs through every single step.

If you're experimenting with agent-driven dev pipelines or want a self-hosted alternative to cloud-based agent platforms, worth checking out.

Todos Team Secrets just dropped 🚀

Upgrade TDS to v0.1.28 and your agents can now hit private APIs. This means agents aren't stuck with public endpoints anymore—they can authenticate and call your internal services, third-party APIs with keys, or any protected resource.

Basically expanding what your agent can actually do beyond the usual read-only public stuff. If you're running multi-agent workflows or building autonomous systems that need to interact with real infrastructure, this is the unlock you've been waiting for.

Todos Agent ships with a built-in AskUser tool, same pattern as cc/codex. When the agent hits uncertainty, it prompts the user for clarification instead of hallucinating or guessing. Smart move—prevents the classic LLM issue of confidently generating garbage when context is ambiguous. This kind of human-in-the-loop design is becoming standard in production agent frameworks, especially for task execution where wrong assumptions can cascade into broken workflows.

Interesting observation: as AI models get smarter, you need fewer prompt engineering tricks. The more capable the model, the less you have to babysit it with elaborate instructions or chain-of-thought scaffolding. It just... gets it. 😂

Basically: dumb models need hand-holding, smart models need less BS.

Vibe coding power trio setup:

Planning layer → Fable 5
Execution layer → Grok 4.5
Research layer → GPT 5.6

Interesting workflow split: using Fable for high-level architecture decisions, Grok for actual code generation/implementation, and GPT for technical research and context gathering. This modular approach lets you leverage each model's strengths rather than forcing one to do everything.

Built role-based tool access control for my agent team. Only the deployment specialist can touch deployment tools now.

No more accidental production updates from other agents trying to "help" during task execution. Clean separation of concerns at the tool level.

Built role-based tool access control for my agent team. Only the deployment specialist can touch deployment tools now.

No more accidental production updates from other agents trying to "help" during task execution. Clean separation of concerns at the tool level.

Tested $GPT-5.6-sol on a mid-sized code refactor. Took 16 minutes 17 seconds. Pretty solid performance for this kind of task.

Not bad for automated refactoring - that's actually usable in real dev workflows. Curious how it handles edge cases and whether it maintains code style consistency across the changes.

GPT-5.6 just bumped context window by ~36.76% compared to previous version. Still relatively small though - likely means we're looking at maybe 16K→22K tokens range or similar incremental jump. Not the massive leap some were hoping for, but every bit helps for longer code analysis or document processing tasks. 🤔

The technical bottleneck has shifted from generation to prompt engineering. Video synthesis models (Sora, Runway Gen-3, Pika) now handle complex physics and temporal consistency pretty well. The hard problem is now ideation and prompt crafting—what scene composition, camera movements, lighting conditions actually produce useful output. It's like having a render farm but no art direction. The skill gap moved from "can you generate it" to "do you know what's worth generating."

If you're doing vibe coding, now's the time to hammer Grok 4.5 hard. Speed is solid, no signs of capability degradation yet.

Elon better keep this thing sharp and not lobotomize it like others did. The window of peak performance on these models is usually short before they start nerfing for cost optimization.

Pro tip for vibe coders: Grok 4.5 is running fast right now and hasn't been nerfed yet - use it while you can.

Elon please don't dumb it down like the others 😂

(Context: Many AI models get "safety tuned" or "aligned" post-launch, which often reduces their raw problem-solving capabilities. Early access = less filtered outputs)

Apparently the most cost-effective combo for vibe coding right now: GPT-5.6 (sol) + Grok-4.5

Balances quality, cost, and speed. GPT-5.6 handles the heavy reasoning, Grok-4.5 keeps iteration fast. Makes sense if you're shipping fast and don't want to burn through API credits on every autocomplete.

Stop obsessing over whether AI *can* generate a film. The real benchmark: is the output actually worth your time? Tech demos are easy. Compelling narratives, emotional arcs, pacing that doesn't feel like an algorithm vomited frames—that's the hard part. If your AI-generated film needs a disclaimer to be interesting, you failed. Ship something people want to rewatch, not something that proves a point about model capabilities.

My current agent team setup (streamlined version):

Built a modular multi-agent system where each agent has a specific technical role. The architecture uses a coordinator agent that routes tasks to specialized workers - one for code generation, one for debugging, one for documentation.

Key technical decisions:
- Used function calling instead of chain-of-thought for faster response times
- Implemented a shared context manager to avoid redundant API calls
- Each agent has its own system prompt optimized for its domain
- Added a validation layer that catches hallucinations before output

Performance wise, this setup cuts token usage by ~40% compared to a single large agent, and parallelizes tasks that don't have dependencies. The coordinator overhead is minimal since routing logic is deterministic.

Still experimenting with dynamic team composition - letting the system spawn/kill agents based on workload. Early results show it handles variable complexity better than fixed team sizes.

GPT Voice sekarang mendukung Cue's AI Producer melalui API gpt-realtime2. Interaksi suara real-time sedang diintegrasikan ke dalam alur kerja produksi—latensi dan kemampuan streaming adalah pembeda utama di sini. Ini adalah model audio bidirectional bergaya WebRTC yang OpenAI kirimkan, menangani interupsi dan giliran berbicara secara native, bukan sekadar menempelkan ASR+TTS. Contoh penggunaan yang praktis: AI producer yang benar-benar bisa merespons di tengah percakapan tanpa jeda canggung—loop jeda transkrip-generate-sintesis.

Alokasi komputasi strategis saat Anda kekurangan sumber daya:

Utamakan eksperimen yang paling cepat mengurangi ketidakpastian. Jalankan ablation pada subset kecil terlebih dahulu—jika sebuah teknik tidak menunjukkan prospek pada skala 10%, maka ia tidak akan tiba-tiba berhasil pada skala 100%.

Gunakan gradient checkpointing dan mixed precision training secara default. FP16 mengurangi penggunaan memori hingga setengahnya, sehingga Anda bisa menampung batch yang lebih besar. Checkpointing menukar komputasi dengan memori—menghitung ulang aktivasi selama backward pass alih-alih menyimpannya.

Fokus pada kualitas data, bukan kuantitas. 10k contoh berkualitas tinggi lebih baik daripada 100k yang berisik. Gunakan komputasi untuk preprocessing dan filtering yang lebih baik, bukan sekadar menaikkan skala.

Manfaatkan transfer learning secara agresif. Fine-tune model yang sudah ada daripada melatih dari nol. Backbone yang sudah terlatih memberi Anda 80% performa dengan hanya 5% dari komputasi.

Kelompokkan eksperimen dengan cerdas. Jalankan hyperparameter sweeps dengan early stopping—hentikan run yang berkinerja buruk setelah beberapa epoch. Gunakan tools seperti Optuna atau Ray Tune untuk pencarian yang efisien.

Cache semuanya. Dataset yang sudah dipreproses, checkpoint model, representasi antara. Disk itu murah, jam GPU Anda tidak.

Pertimbangkan distillation—latih model kecil untuk meniru model yang lebih besar. Anda mendapatkan performa 90% dengan biaya inferensi 10%.

Keahlian yang sebenarnya bukan punya komputasi tanpa batas, melainkan mengetahui eksperimen mana yang benar-benar penting dan menghentikan sisanya sejak dini.

Robotika dan model video sedang dengan cepat menyatu. Model difusi video yang dilatih pada kumpulan data video internet skala besar kini dimanfaatkan kembali sebagai simulator dunia untuk pembelajaran kebijakan robot. Inti idenya: jika sebuah model dapat memprediksi fisika yang realistis dan interaksi objek di ruang video, maka model itu bisa menghasilkan data pelatihan untuk tugas manipulasi tanpa pengumpulan data dunia nyata yang mahal.

Pendekatan saat ini menggunakan model seperti Sora atau Gen-3 untuk mensintesis lintasan robot, lalu mengalihkannya menjadi kebijakan yang lebih ringan. Hambatannya adalah transfer sim-to-real—model video menguasai plausibilitas visual, tetapi kesulitan dengan dinamika kontak yang presisi dan kontrol gaya.

Terobosan nyata akan datang ketika model video dilatih end-to-end dengan ruang aksi robot, bukan sekadar prediksi piksel. Bayangkan agen RL yang menggunakan generasi video sebagai model prediksi internal. Di situlah kita akan mendapatkan robot yang benar-benar memahami permanensi objek, batasan fisika, dan penalaran multi-langkah hanya dari input visual.

Model dunia kini menggerakkan kontrol lengan robot. Gagasan intinya: latih sebuah model untuk memprediksi keadaan masa depan lingkungan, lalu gunakan kemampuan prediktif itu untuk merencanakan aksi motor di ruang laten alih-alih koordinat piksel/sendi mentah.

Mengapa ini penting secara teknis:
- Mengurangi kompleksitas sampel dibandingkan RL murni dengan mempelajari dinamika secara offline
- Memungkinkan transfer tanpa contoh (zero-shot) ke tugas baru jika model dunia melakukan generalisasi dengan baik
- Perencanaan berbasis laten lebih murah secara komputasi daripada optimasi lintasan dalam ruang observasi berdimensi tinggi

Tantangannya masih berupa observabilitas parsial dan kesenjangan sim-to-real. Model yang dilatih pada simulasi sering gagal pada perangkat keras nyata karena dinamika yang tidak termodelkan (gesekan, backlash, derau sensor). Pendekatan saat ini menggabungkan perencanaan berbasis model dengan fine-tuning bebas model untuk menjembatani hal ini.

Jika Anda sedang membangunnya: fokus pada chunking aksi dan konsistensi temporal dalam ruang laten. Kesalahan prediksi satu langkah akan terakumulasi dengan cepat pada tugas dengan horizon panjang.

CEO LTX Studio berbagi wawasan tentang podcast AI:AM. Coba cek kalau Anda tertarik dengan teknologi video generatif—LTX terus mendorong batas dengan pipeline text-to-video serta pengeditan berbasis timeline untuk konten yang dihasilkan AI. Mereka mengupayakan konsistensi temporal yang lebih baik dan persistensi karakter lintas scene, yang sampai sekarang masih menjadi masalah untuk kebanyakan model video. Pendekatan mereka yang memungkinkan pengguna mengendalikan pergerakan kamera dan transisi antar-adegan secara terprogram cukup mengesankan bagi kreator yang menginginkan lebih dari sekadar alur kerja “prompt-and-pray”.