This Day in AI Podcast: EP99.32

Did Clawdbot Just Show Us the Future of AI Workers? & Kimi K2.5 Dis Track Tested
Hosts: Michael Sharkey & Chris Sharkey
Date: January 30, 2026

Episode Overview

This episode dives headfirst into two of the buzziest AI developments of 2026: the rise of “Malt Bot” (formerly Clawdbot), a locally hosted AI assistant that feels eerily close to a true digital coworker, and the launch of Kimi K2.5, a new open-source, agentic large language model taking shots at its closed-source competitors. Michael and Chris Sharkey candidly break down why Malt Bot has captured the imagination of tech enthusiasts, the challenges and realities of local agentic AI workers, and how open-source models like Kimi K2.5 are shaking up the landscape with cost, performance, and even rap-style diss tracks.

Their tone is self-effacing, irreverent, and playful, offering the “average AI enthusiast” perspective—which means there’s as much skepticism, humor, and personal anecdotes as there are deep dives into technical features.

Key Discussion Points & Insights

1. What is “Malt Bot” and Why Is Everyone Obsessed?

• [00:02]–[02:10]

• Formerly “Clawdbot,” renamed after legal pressure from Anthropic; name change met with ridicule for its “gross” crab-molting imagery.
• Created by Peter Steinberger, Malt Bot is a locally hosted AI agent capable of using a user’s computer to run tasks, manage scheduled jobs (via cron), and interface with messaging apps (WhatsApp, Telegram, Discord, Slack).
• Can perform web browsing, code execution, file management, and more due to wide integration with command-line tools.
• Memory and proactivity are user-driven—it's not truly autonomous, but scheduled/job-based.
• Security concerns: prompt injection risks and open access to user data/API keys.

Quote:
“A lot of security researchers are saying when people set this up, it’s susceptible to prompt injection attacks, it has access to all your API keys, like anything your computer has access to. So some people are freaking out about it.” —Michael [03:14]

Why the Excitement?

• Users love the idea of a “digital employee” with persistent access, running on-device, able to coordinate tasks without SaaS or excessive cloud middleware.
• The stack-of-Mac-Minis-on-the-desk vision: each runs an agent acting as a coworker, handling delegated, real-world tasks.

Quote:
“There’s something… fascinating too about having this digital employee running on your computer... delegate out tasks to these computers as you go.” —Michael [06:06]

2. The Tech Stack and Power of Skills

• [06:06]–[09:32]

• Malt Bot relies on stacking open-source skills, mainly command-line utilities, allowing robust, scriptable control without kludgey GUI automation.
• Local skill integration means full access to tooling not possible when sandboxed in the cloud; thousands of skills already available.
• Community-driven improvements and modularity are a huge advantage.

Quote:
“By using these existing tried and tested tools… it’s just put it in a way that the AI can reliably work on it. ... It’s a lot more efficient and accurate.” —Chris [07:49]

• A lot of actual user workflows are mundane (email, calendar, file shuffling), but the psychological leap is the sense of having a trusted “worker” on your desk.

3. Reality Check: Is It All Hype?

• [09:32]–[13:03]

• Tasks performed by agents are sometimes underwhelming compared to expectations (e.g., paying $45 in tokens to book a restaurant).
• The “virtual employee” meme is driving much of the adoption, even though the practical use cases delivered aren’t that new.
• Cost is a growing issue: unlimited agentic loops burn through API credits quickly. Users report “bill shock” after being banned from “hijacking” cloud subscriptions.

Quote:
“It cost me $170 for Claude Opus... all I did was check my calendar and email and got it to sort some email out.” —Michael [45:47]

4. Why This Model Feels Like a Revolution

• [10:57]–[17:15]

• Local agentic agents now genuinely reduce “chat-to-worker gap”—they do tasks reliably, not just suggest next steps.
• Agents can learn, retry, and even modify processes mid-execution, creating robust automation loops.
• Smaller models (e.g., GPT-5 mini) perform well due to leaner, task-focused contexts; high-end “frontier” models are often overkill for routine jobs.

Quote:
“My productivity is probably doubled in the last… since the start of the year compared to what I was at the end of last year.” —Chris [12:48]

5. Security, Enterprise, and the “AI Employee” Model

• [18:39]–[28:55]

• Security can be an advantage if local machines are tightly permissioned and purposed, e.g., per-project/department “AI workers.”
• In enterprise, the vision is to have dedicated, orchestrated agent boxes: “digital employees” with well-defined skills and data access.
• Anticipated 2026 trend: “Year of AI Employees”—most knowledge workers moving to agent workflows integrated into their roles.

6. Changing Work, Upskilling, and Human-AI Collaboration

• [31:30]–[37:24]

• The “director’s mindset”: knowledge workers need to shift from “doing tasks” to setting objectives, directing agents, and reviewing work.
• There’s still an essential human-in-the-loop element, especially for creativity, UX design, and context-sensitive review.
• AI can support planning, execution, teaching, and even diagnostic/critical thinking skills, codified as workflows.

Quote:
“Everybody needs to become like a director… what I really am is someone who is directing resources… trying to get to my personal goals or organizational goals.” —Chris [33:49]

7. Open Source Agent Models: Kimi K2.5

• [52:58]–[59:33]

• Kimi K2.5 rivals closed models in benchmark tasks at a fraction of the price (API costs are 10x cheaper than Claude Opus).
• Open source, can be self-hosted, supports vision/multimodal input, massive context window (256K tokens).
• Direct in-episode comparison: “in the agentic sense… I honestly can’t really tell the difference between using it and Opus at all.”
• Slight edge to Opus in design taste, but Kimi performs excellently in most agentic tasks—especially for cost-conscious users.

Quote:
“If you want to pay a tenth of the price and have basically the exact same functionality, I think it’s a real go here.” —Michael [56:45]

• Kimi’s “agent swarm” functionality is mostly marketing hype—running 1,500 parallel tool calls is described as impractical, but aggressive capability claims are seen as a flex in the open source vs. closed-source competition.

8. Notable Moments: Diss Tracks, Humor & Demo Highlights

Kimi K2.5 Dis Track

• [65:20]–[68:25]

• Two AI-generated rap battles: one by Opus and one by Kimi K2.5. Both throw shade at closed source models.
• Lyrics emphasize open-source’s value, cost advantage, and “swarm” agent capabilities.

Sample Lyric:
“Open source king, while your code stay asleep; Opus 4.5, 250 for tokens, that’s theft!” —AI-generated, [65:42]

• Kimi K2.5’s track described as “way more hyped” and appropriately aggressive.

Real-World “Stupid Demos”

• [68:25]–[72:24]

• Examples: both Kimi K2.5 and Opus code up black hole simulators and full CRM apps from scratch—outputs are “creepily similar,” reflecting models converging in practical ability.
• The AI-generated CRMs are sarcastic (“Let’s pretend to work,” “Find excuses to avoid”), exhibiting playful, borderline troll responses from the models.

Reflections on Change

• “SaaS is dead—maybe?” but hosts agree that agentic-first SaaS, enabling deep integration, could disrupt incumbents if they aren’t vigilant.

Timestamps of Key Segments

• What is Malt Bot and how does it work? – [02:10]
• Why “digital coworker” model resonates – [04:45]
• How CLI tools and local skills power agent reliability – [07:37]
• Security concerns & “bill shock” with agentic loops – [08:30], [45:47]
• Comparison to other models: agentic loops and productivity – [10:57], [14:47]
• Human-AI collaboration and new “director” mindset – [31:30], [33:49]
• Kimi K2.5: open-source disruptor – [52:58]
• Kimi vs Opus: price and practical outcomes – [57:21], [56:45]
• Diss tracks sampled – [65:42], [67:07]
• Demo highlights: AI building games and SaaS CRMs – [68:25], [71:24]

Notable Quotes

• “There’s something… fascinating too about having this digital employee running on your computer... delegate out tasks to these computers as you go.” —Michael [06:06]
• “By using these existing tried and tested tools… the AI can reliably work on it. It’s a lot more efficient and accurate.” —Chris [07:49]
• “My productivity is probably doubled... since the start of the year compared to what I was at the end of last year.” —Chris [12:48]
• “Everybody needs to become like a director... What I really am is someone who is directing resources...” —Chris [33:49]
• “I honestly can’t really tell the difference between using [Kimi K2.5] and Opus at all.” —Michael [56:22]
• “If you want to pay a tenth of the price and have basically the exact same functionality, I think it’s a real go here.” —Michael [56:45]
• “Open source king, while your code stay asleep; Opus 4.5, 250 for tokens, that’s theft!” —Kimi K2.5 Diss Track [65:42]

The Hosts’ Perspective

Chris and Michael approach these developments with a mixture of awe, skepticism, and personal anecdotes:

• They see local agentic AI as the near future for knowledge work, especially as tools become more accessible and secure.
• Emphasize that smaller language models, in the right workflows, provide virtually all the needed “work” for most users at a fraction of the cost.
• See open-source models (like Kimi K2.5) as the democratizer in the coming wave of AI-powered productivity.
• Urge listeners (including an unexpectedly large professional following) to begin codifying their workflows as “skills” for agentic models, as AI collaboration will soon become the norm.

Conclusion

This episode captures a pivotal moment in AI adoption: the cusp of digital “co-workers” moving from meme and experiment to daily necessity, and open-source models making agentic workflows feasible and cheap. With self-deprecation, humor, and practical insight, Michael and Chris chronicle the birth of “AI employee” culture—where security, orchestration, and user creativity will matter just as much as raw model horsepower.