#01 - Oct 16, 2026
State of AI: Help Saving Harmony
INSEAD's AI Native study, Google on self-learning, Savings on your LLM bill
Read plain text
Anthropic’s Scion https://www.anthropic.com/engineering/redesigning-source-control(piper https://en.wikipedia.org/wiki/Piper_(source_control_system) or sapling https://sapling-scm.com/ much?), Gemini 4 https://x.com/rauchg/status/2110044152482365747?s=20, Muse Wonder https://x.com/theo/status/2110271073581421003?s=20, Astra debacle https://x.com/theo/status/2110286054460883106?s=20, Embedded https://huggingface.co/spaces/webml-community/embeddinggemma-2-webgpuGemma2 https://eg2.haku.sh/, are things you already read. Here’s what really happened. ---------------------------------------- Get Help tl;dr: Surprise, Surprise! Industry knowledge sharing is a top indicator of success. INSEAD https://papers.ssrn.com/sol3/papers.cfm?abstract_id=6513481 has run a shadow experiment with great results for those AI adopters: treated firms complete 12% more tasks, are 18% more likely to acquire paying customers, increased +1.9x revenue, +44% AI use cases, and ~40% capital req. Reduction. Plus better conversion rates, and delivery acceleration: * Everyone is born equal The control group had the same tools, the same credits, and the same technical training as treated ones. Also, they all had 1-hour introduction, 6x90-minute deep dives, weekly peer learning & office hours. All optional. So what was the difference? * Do you Hear or Listen? The binding constraint on AI's value at the firm-level is not access to the technology but search over where to deploy it. Firms that reorganized internal workflows by replacing sequential, labor-intensive processes with AI benefit the most. That means, they changed the processes, they didn’t simply accelerate the steps (think faster cars vs a plane). A field services venture compressed five roles into AI modules that self-improve from operator feedback. Hence, the highest-value applications that require rethinking how work is organized went undiscovered without an external input. * 10 weeks / 500+ startups. The ten-week experimental window captures short-run effects only. Whether these performance gains compound or dissipate as ventures mature remains an open question. Early-stage firms have low organizational inertia, when organizations grow, hierarchies deepen, processes calcify, and decision authority becomes distributed. If this heterogeneity is present even among early-stage ventures with minimal organizational complexity, the challenge is likely greater still for established firms. → Adopt. The bottleneck is not the Tech team, it is the managerial challenge of defining an executive vision, discovering where the technology creates value, and giving a clear mandate to the organization. You are not alone. ---------------------------------------- Harnessing Harmony tl;dr: Google on how they discovered self-evolving harnesses https://arxiv.org/abs/2610.15150. A very interesting paper for those of you implementing your own harnesses, here the learnings: * New approach. Jev is extremely cheap, however labs are open sourcing their models as response. Kind-of what’s happening to some company brains https://supermemory.ai/blog/open-sourcing-company-brain/, or even DevinAI https://x.com/cognition/status/2107165034463867001?s=20. People are getting their hands forced into OSS. This is the shortest race to the bottom we’ve seen mostly because... * Old smell. They propose several of the basics: dreaming, lazy loading of skills and memories, self modifying agents. The concepts you are familiar with, the key insight is the rankings that they use. They don’t have evolution gates, as many do, but they are grading them. We do the same. * Shortsighted They evolved their systems ~4 times, and the benchmark has 174 tasks. For comparison, agent-swarm.dev ran 50000+ tasks in production this past month, and nightly evolution for 225+ nights. This is without even mentioning the 45 models we use, vs Gemini Flash. Large companies are lagging behind and they don’t consider cost. It's not just Google. Devin just launched an oncall https://x.com/cognition/status/2110790097369010292?s=20 product. Every team we worked with, implemented this, months ago (blueprint https://docs.agent-swarm.dev/docs/playbooks/open-oncall). It was the first painful thing to do. Ask yourself, why took them this long? → Wait. Just remember, you are on the right track, if 6 months ago you were here. If you are not, then a big corp just moves faster than you. Ask for help. ---------------------------------------- Saving the Maestro tl;dr: Update your swarm pronto! We’ve been heads down optimizing costs, here the goodies. This past week we’ve been experimenting with optimizations, this is what we found out: * Leading is expensive Remember this https://www.pleasedontdeploy.com/p/our-strategy-to-deal-with-llms-prices? Well basically since then we have been working hard to reduce the % of spending that goes to the Lead vs other agents. We are getting closer to 50/50. Yes, we started using Opus 5.5, it’s great, but... * System One Beats We found out two simple ways to optimize for no-ops and cache use. The first one, optimized classification with Laya/Cleff/Jev. When applied this to workflows, heartbeats, and some other special tasks it works better than inference at a 1/40th of the cost. The second one: prompt magic. If you are using Claude in particular, we are fine-tuning the prompt order to further optimize for caching. * Secret sauce. We are working on an engine to optimize the swarm specifically for each setup. You’ll soon have an extension that will cut your bill in half 🤞. If you want to be in the early alpha, let us know. In the past week we tried different open weight models, and subscriptions, with the feeling that this could be done better. Still we don’t have our own GPUs. → Challenge. Make sure you keep the swarm updated, or you watch the repo frequently. Check the code, see what's done, and make it better. ---------------------------------------- Go update your swarm, start saving monies https://github.com/desplega-ai/agent-swarm, forward this email to your bestie, Or go read slop https://x.com/GergelyOrosz/status/2111055157676986516/photo/1. Your call. Best, ---------------------------------------- Sent with ❤ by desplega.sh http://desplega.sh/ unsubscribe .