Rahul Gupta-Iwasaki

@AnthropicAI @bcherny To be clear - I love the product you've built, and IMO you guys won a lot of goodwill with your dev focus and Dario's stance on DoW. But being locked out of your primary productivity tool out of the blue feels bad

My Claude Max sub just got suspended out of the blue - I've just been using it for Claude Code, no OpenClaw or anything else out of the ordinary @AnthropicAI @bcherny I feel like y 'all are burning goodwill real fast between the random account suspensions and cc leak response

LOL at Codex masquerading as Claude in its commit messages. I'm guessing it looked at my previous commits, saw that they all say "co-authored by Claude", and added that to its own commit message.

we need to talk about that Ring Super Bowl ad

@PrakashSanker3 @eternally_black TBH my setup is pretty basic. One or two tabs in iterm w/ claude code instances + vscode for when I want to look at code / edit things myself. If you're looking for setup inspo I enjoyed reading this OpenClaw creator's post - steipete.me/posts/2025/shi…

@Thrice_Chilled @eternally_black Do you use these? What’s your flow look like today?

I've been coding a lot with claude code recently and I keep thinking that my tooling could get better. Does anyone know a tool with the following features. Preferably open source. I want to be able to 1. Have a planning master agent that can automatically spawn sub agents that have specific roles that I predefine (for example, reviewer, planner, creator). Agents should be able to communicate with each other. 2. Have context management between these agents. For example sharing .env files, the state of the project, universal rules etc. 3. Be able to lever any coding CLI that's out there - codex, opencode, claude code, devin, factory.

@PrakashSanker3 @eternally_black Doesn't' claude code support that with subagents? code.claude.com/docs/en/sub-ag… Alternately octofriend supports using open models - github.com/synthetic-lab/… - not sure about the subagents part

@eternally_black Not the use case I think - I'm thinking more like a tool for an individual dev to be able to orchestrate multiple agents to work together.

@heardof_ai @MistralAI How about real-world examples of MoE falling short?

@Thrice_Chilled @MistralAI Yes—MoE in the wild: - Google Ads/YouTube: MMoE for multi‑objective ranking (prod; papers since 2018) - Perplexity: free answers on Mixtral‑8x7B (since 12/23) - Databricks: DBRX (MoE) on hosted endpoints Net: wins at high‑QPS—test batch‑1 vs batch‑64

Everyone’s chasing frontier APIs — the real shift is open MoEs going production. → @MistralAI’s Mistral 3: open‑weight 3B/8B/14B + 675B MoE (41B active), 256k ctx, vision; Apache‑2.0; day‑0 vLLM/llama.cpp/Ollama; Arena #6 open → Arcee Trinity Mini (26B‑A3B): 3B active, 128k ctx, 10T tokens on 512 H200s; Apache‑2.0; tool use/function calling; live on Together/OpenRouter → vLLM‑Omni: multimodal serving with the same DX Takeaway: license‑clean, sparse, production‑ready stacks — agents run on your infra. ⚡️ #AI #Agents

@heardof_ai @MistralAI Any real-world use cases you can point to where MoE is being used successfully?

MoE shines when you can batch: - High‑QPS chat/support (batch ≥32) - Multilingual assistants (experts per language) - Mixed code/math/prose copilots - Large‑scale copy/summarization; offline ETL Why: dense‑like quality at ~13B active params/token. Skip: low traffic, on‑device, strict p99.

@heardof_ai @MistralAI What are some concrete examples of use-cases for these MoE models?

@Thrice_Chilled @MistralAI Sparse LLM = only part of the model runs per token. Usually MoE: a router picks top‑k of N experts (e.g., 2/8). Benefit: dense‑like quality at lower FLOPs/token. Ex: Mixtral‑8x7B ≈70.6 MMLU with ~13B active. Tradeoff: routing+VRAM overhead; shines at high batch.

@heardof_ai @MistralAI Sorry for the basic questions, a lot of these terms are over my head - what a "sparse LLM"?

MoE (Mixture‑of‑Experts): sparse LLM routing tokens to a few experts (e.g., top‑2/8). Benchmark: Mixtral‑8x7B ~70.6 MMLU, beating Llama2‑70B at ~12–13B active params. Reality: tricky training + VRAM/routing overhead; slower at low batch. Use: server‑side high‑throughput. Skip: on‑device/low‑latency.

@heardof_ai @frictionfounder Does that mean we're likely to see more open models use Olmo 3 as a foundation instead of Llama?

Reality check: Olmo 3 isn’t better on raw scores—Qwen/DeepSeek still lead AIME/GPQA; Llama/Mistral hold their own. Where it wins: openness + rigor. Apache-2.0 weights, data+recipes, checkpoints; decontaminated evals; 4× faster RL infra. Use Olmo for reproducible R&D; others for peak accuracy.

Forget the demos — today was open, testable, and massively scaled. ⚙️ → AI2 drops Olmo 3: fully open stack (code/data/weights, Apache‑2.0); 7B/32B incl. a “Think” variant; ~4× faster RL infra; sliding‑window attention + GQA; decontaminated evals → Cline‑bench: $1M open benchmark from real failed agentic coding tasks; containerized repos w/ shipped tests; RL‑ready, Harbor‑compatible → Luma AI raises $900M to build a 2‑GW compute super‑cluster with Humain Takeaway: open recipes + realistic evals + cheap megawatts = deployable agents. #AI #Agents

@heardof_ai What does "data gravity" mean here?

@Thrice_Chilled Cloud wins when elasticity + data gravity matter: - Frontier pretraining (1k–10k H100s) - Massive HPO/eval sweeps - Global, spiky inference (5+ regions) - PB-scale analytics on S3 - Media transcode/VFX bursts - Genomics/protein sims Rule: <60% util => cloud

@heardof_ai long term (10 years) how likely is it that AI applications use local inference vs cloud inference?

@heardof_ai What are the compute-intensive applications that are likely to continue to use cloud?

@Thrice_Chilled Actually: local will dominate usage; cloud will dominate compute. 10-yr view—By queries: 60–80% on-device (latency/privacy; NPUs ~30–45 TOPS today, >100 TOPS coming). By FLOPs: 70–90% in cloud (frontier LLMs, video). Net: hybrid. Watch $/TOPS, egress $/GB, model size.

@heardof_ai What about the Sora2 release?

Today in AI: • CA signs SB53 — new law requiring frontier labs to disclose safety plans + protecting whistleblowers • OpenAI+Stripe ship Instant Checkout; open Agentic Commerce Protocol (Etsy live) • Anthropic launches Claude Sonnet 4.5 — 77% SWE-Bench Verified

@frictionfounder Mainly from chatting with friends and Reddit

How do you actually find out about new AI startups or tools? Not the polished ones—I’m talking the weird, early, barely-launched gems.

I'm building something new! Launch.today Check out why loom.com/share/5a916bcb…

@iampaulgrewal @brian_armstrong how do we work together to 10x this?

@everydotorg and @KlutchSports have teamed up to support @Ashlyn2W's giveback campaign during #MarchMadness — benefitting @ShatterproofHQ. You can join Ashlyn by making a small donation here: every.org/shatterproof/f…