Alignment is subjective. Truth is objective.

Major UX boost! 🙏

Until a week ago, @Grok Build was running very short response loops which made it tiring and mundane to babysit one session, let alone multiple. Last few days, response loops run much longer and do a lot of work before stopping for user feedback. It now feels more productive and satisfying to work on multiple projects simultaneously. The better AI gets, the more fun it is to take on bigger challenges.

Note to self and founders. Enterprise SaaS UI was the biggest monster created by the software industry. Proprietary data is gold. AI agents should be able to use it and turn into actionable recommendations. But there is little room in the future for siloed UI apps for humans. x.com/theallinpod/status/206…

Snapshot of open source timeseries models: - Scaling laws work for timeseries transformer based models - Specilist models outperform generalist models at a fraction (1/1000) of trained model parameters (<100M-1B vs ~ 100B-1T ). - Timeseries fine tuning on domain specific data still matters. No zero-shot champion across all kinds of time series.

Notable update. 1. Scaling laws work for time series. 2. Specialized models can outperform large generalist models with a fraction of trained parameters.

A couple of observations after months of hands-on tinkering with OpenClaw, Hermes, and Nanobot. We are not there yet. None of the frontier models have reached a point where they can see the whole picture in a real world company. I’ve repeatedly tried to use a single control board for planning, building, and consuming tasks. This works at small scale with a few tasks, skills, and cron jobs but falls apart after about 10 unrelated tasks. The harness and model lose coherency, intermingling instructions from unrelated skills, running code built for other tasks, and creating an environment mess that takes nontrivial effort to unwind. Harness upgrades continue to be a major friction point. Each upgrade leads to hours of debugging broken configs, isolation, access levels, repo states, and corrupt backup scripts. Agent-to-agent communication is far from working. We are several iterations from establishing meaningful, reliable, and smooth protocols. At this stage, agents negotiating with each other remains an art that quickly gets out of hand and needs frequent supervision. The good news is there may be a practical middle ground. Code-building harness tools are improving at handling longer, more complex tasks and are tuned for bigger projects. @Grok Build is catching up fast to Claude Code and Codex. The pattern that works in practice at this early stage of the agentic AI shift is as follows: Agentic AI Front End: - Consumer harness for the billions of non-builder users: memory and lightweight skills capturing personal preferences, focused on personal UX style and taste. - Analogous to how consumers pick food, clothes, and furniture: end users don’t make these things but have strong preferences. Agentic AI Backend: - Reliable, battle-tested MCP servers with utility-grade reliability and uptime. - About one in a thousand people are builders. Millions of builders (versus billions of users) need professional tools. - The agentic AI backend needs to be utility-grade, predictable, and reliable like water, electricity, roads, internet, or blockchain. - Consumers don’t care how backend services are built but get upset when utilities are unavailable or subpar. Overall recommendation for founders: - Focus on building agentic-friendly commercial MCP services using solid code-building tools. - Don’t spend too much time on human-facing monolithic web and mobile apps. - Publish a few example front-end skills that end users can feed to their favorite chatbots, understand, customize, and add as lightweight SKILL Connector packages. Let the chatbot render UX in the user’s preferred style.

Does anyone outside Amazon have the ability to verify the capabilities of this new bot? I'm getting fatigued from robotics AI announcements and cool videos that are not independently verifiable.

It’s been several weeks and still no sign when @grok voice will integrate with connectors and skills in Grok App? On mobile , web and @Tesla.

Significant milestone. Transparency to what Grok actually remembers and ability to edit (i.e. label and help Grok train its learning model to personalize memories better). This may be one of the hardest frontier tasks, because people vary widely in their preferences. Same problem with Tesla FSD parking. Still not solved. It's not about safety (everyone agrees), but personal preferences (everyone has their own taste and style ... that change with time, mood, weather and emotional state).

Kudos to @RobinhoodApp for offering Agentic MCP for market data and portfolio actions. Works great with Grok App Connectors. robinhood.com/agentic @SchwabTrading , @Fidelity , paying attention?

Highly valuable long form podcast on history and mechanics of perps.

Grok Build noticeably improved in the last few days. Thinks harder, works longer, makes fewer mistakes on non-trivial projects. Maybe most notably: fewer repeat mistakes. Once corrected, it seems to try harder to avoid regressing.

Stand with Crypto. Reach out to your senator. Or do nothing and let legacy too big to fail banks keep running government.

Robotics models making progress. Not quite as aggressive as LLMs but progress nonetheless:

Good to see @resend connector in @grok Build. Way easier for agentic flows than gmail. Next stop , Grok App connector for resend? Yes, with webhooks support in Grok. Let’s go!

Grok coding is still not too impressive but in a strange way recently its ability to use skills in @grok app , do deep grounded data analysis , plot annotated charts and reason probabilistically over forecasts is now at a whole new level. It seems like the world is bigger than coding.