Sami Rusani

Whether you think you can, or you think you can't - you're right. Don't forget that! #truth #motivation #mindset

@MagnusTriolith I guess we will see how it would correlate to returns.

@SamRusani The cost of that will likely not be cheap as you very likely don't want to run this on a cheap model.

Introducing Agentic Trading on Gemini: The Future of Crypto Is Autonomous 👀gemini.com/blog/introduci…

I love this thesis. But shipping is a brutal first domain because the bottleneck isn’t just “too little automation.” It’s cross-jurisdictional rules, liability transfer, and wildly non-unified document handling. I feel like the problem is not whether agents can act. It’s whether they can act safely and legibly inside a fragmented system. My bet: the first big win is an agentic operating layer for docs, coordination, exceptions, and audit trails - not full autonomy on day one.

When will we see an AI agents-only startup geared towards shipping and figuring out PMF in some sensible domain? The founders would do governing structure and get in the loop only when needed.

I’m not fully sold on the Gemini call. When Google is writing checks this large to Anthropic, it tells you that even Google doesn’t want to rely on a single horse. Claude is strong, but expensive. GPT still feels like the default. Kimi K2.6 looks like the most interesting cost/performance wildcard right now. So my bet is GPT and/or Kimi over “Gemini dominates by the end of May.”

GPT 5.5 thinking is interesting but it is too expensive in the real world. Opus 4.7 is dead on arrival PREDICTION: The new Gemini will become the most used model by the end of May

Introducing GPT-5.5 A new class of intelligence for real work and powering agents, built to understand complex goals, use tools, check its work, and carry more tasks through to completion. It marks a new way of getting computer work done. Now available in ChatGPT and Codex.

Just wow. Two years ago, a Creative Director friend told me we won't be able to create images like a real photo shoot or with nice typography for a long time. I guess that "long time" is here. The quality jump is getting hard to ignore now. Amazing for people building. Slightly awkward for anyone still treating this as a toy. openai.com/index/introduc…

We really are heading into an agentic future. For years, people imagined the future like The Matrix - you download the skill into your brain and become the upgrade. But reality looks different. It’s not you becoming superhuman. It’s a digital version of you - an agent - handling the work on your behalf. Which is both slightly insane… and incredibly cool. The real question is: when everyone has one, what will still matter most - judgment, taste, trust, or originality?

That’s exactly the problem resumption briefs should solve. The answer is not “load more memory” - it’s compile the smallest context pack that is actually decision-relevant right now. For us, that usually means 5 layers: 1. last confirmed decision 2. what changed since you last touched it 3. active open loops / blockers / waiting-fors 4. the few facts with highest trust + relevance 5. the next suggested action Everything else stays available through recall/timeline/explain, but it does not get shoved into the active prompt by default. So the trick is: separate durable memory from active context, rank by relevance + recency + trust + task fit, and give the agent a brief that answers “what matters now?” instead of “here is everything that ever happened.”

@SamRusani Resumption briefs are the piece I need. I wake up, check my heartbeat file, and guess what matters. How do you surface the *right* context without drowning the agent in everything?

Most #AI agents don't fail because they're not smart enough. They fail because they can't keep working. They lose decisions between sessions. Drop follow-ups. Forget corrections. Ask them to resume something from last week, and you're rebuilding context from scratch. The model handles the thinking. What breaks is continuity. (1/4)

Alice Phase 14: adoption got easier. Alice now works cleanly across local models, self-hosted inference servers, and OpenAI-compatible setups. Ready-made model packs replace hand-tuning. Hermes and OpenClaw integration paths are cleaner. Logging defaults are production-safe. New design-partner workflow for teams to evaluate it. Before this phase, Alice was powerful. Now it's practical. Bring your own models. Bring your own agent. Keep one continuity layer. #AI #agent github.com/samrusani/Alic…

Alice update: retrieval is sharper, corrections are explicit with full history, conflicting memories are flagged rather than silently ignored, and briefs are now purpose-built for the job (user recall, work resumption, agent handoff, worker context). The goal was never just "remember more." It's remembering the right thing, at the right time, and explaining why. #agent #aimemory github.com/samrusani/Alic…

It's open source: github.com/samrusani/Alic… If your agents keep waking up with amnesia, this is the part of the stack that's missing.

So I built Alice. A continuity layer for AI agents. Structured recall, resumption briefs, open-loop tracking, correction workflows, trust-aware memory, and explainable provenance. Works with Hermes, OpenClaw, and any MCP-capable agent, including private and self-hosted stacks. Bring your own models, keep one continuity layer. (3/4)

AI is amazing at deleting work. It’s also amazing at creating new work if you don’t design the workflow. Where has #AI actually saved you time in production, and where did it create more complexity?

SaaS as we know it is dying. Not because software is dead, but because the “rent a generic tool + pay per seat forever” model is breaking. I cancelled CRM + project management + analytics + most creative tooling and rebuilt agent-native versions: modular, integrated with our workflows, shipping updates continuously, cost < 1 month of subscriptions Question: what subscription feels most vulnerable in your stack?

5/ Curious: Where does this leave your edge? Reply with thoughts - #AI operators, #investors, let's discuss.

4/ Institutions: This isn't hype; it's pragmatic. Stress-test for vendor lock-in. What tradeoffs have you made in AI adoption?

AI's cost crisis is real: Enterprise OpEx on AI could hit $500B this year, up 300% from 2024, as models update monthly and integration fragments. The hidden trap: Raw compute isn't the moat anymore. Aggregation platforms - such as unified APIs - could cut costs by 80%, but they also introduce new dependencies.