🏭 "You Are Not Hireable" and Other Bold Claims
Field notes from the AI trenches—what actually matters this week
This week, Sequoia Capital declared AGI has arrived - their definition, their timeline. They’re pointing at systems that already work autonomously for 30 minutes straight, projecting a full day’s work by 2028. Meanwhile, a grassroots movement of developers is buying Mac Minis to run 24/7 AI assistants that code, automate, and extend themselves. One early adopter’s take: “If you don’t have at least 3 long-running agents, you are not hireable.” Hyperbole? Maybe. But they’re betting real money on it.
The productivity divide isn’t coming - at least according to early adopters. $599 down, $100-200/month in API costs, and you get a digital employee that works while you sleep. Whether this becomes a real advantage or just expensive tinkering remains to be seen. But the people using it aren’t hedging: one developer says it “feels like cheating”, another claims “if you don’t have at least 3 long-running agents, you are not hireable”. On that last claim, time will tell.
This week’s newsletter: long-horizon agents hitting prime time, the Clawdbot revolution, Claude’s philosophy lesson, the ecosystem building agent superpowers, and the uncomfortable moment the White House started manipulating arrest photos with AI.
🚀 Sequoia Says AGI Arrived on January 14th
What happened
Sequoia Capital declared 2026 as the year AGI arrives through long-horizon agents. Not AGI as science fiction, but AGI defined functionally: systems that can figure things out through sustained autonomous work.
What it does
Agents work reliably for ~30 minutes today, with performance doubling every ~7 months
Already functioning as specialists in medicine (OpenEvidence), law (Harvey), cybersecurity (XBOW), and DevOps (Traversal)
Projected timeline: full day’s work by 2028, full year by 2034, century by 2037
Why you should care
Sequoia isn’t philosophising about consciousness - they’re citing METR data showing these systems doubling in capability every 7 months. A century of compute means 200,000 clinical trials finally cross-referenced, every customer support ticket mined for patterns no human would spot. Their projection: 2 years from agents that work a full day, 8 years from agents that work a full year. Whether these timelines hold is anyone’s guess, but VCs are placing bets accordingly.
The stakes (if Sequoia’s right)
If these projections hold, the question shifts from “will AI take your job” to “are you the one wielding century-scale agents or competing against them.” Big if. But worth watching.
🏠 Clawdbot: The $599 AI Employee That Codes From Your Phone
What happened
Federico Viticci published a deep dive into Clawdbot, an open-source AI assistant that runs locally, stores everything as Markdown, and can improve itself. He burned through 180 million API tokens in a week of experimentation - depending on the input/output split, that’s probably burning $1,200 - $2,000+ in one week of heavy experimentation. That’s when a cost-capped subscription plan becomes important.
What it does
Runs as local agent with gateway to iMessage, Telegram, WhatsApp
Executes Terminal commands, writes and runs scripts, installs skills, sets up integrations
Self-improving: can give itself new features and capabilities
Integrates with Notion, Todoist, Spotify, Sonos, Philips Hue, Gmail, and more
Via the “coding-agent” skill, it can spawn and control Claude Code sessions programmatically
The Mac Mini trend
A grassroots movement on Twitter: people buying base Mac Minis specifically to run Clawdbot as dedicated home AI servers. One user: “buying a mac mini for @clawdbot is like getting a hermit crab a new shell.” The $599 M4 Mac Mini becomes an always-on AI assistant that controls smart home devices, manages calendars, and executes complex automations via any messaging app.
The remote coding angle
Developers are using Mac Mini + Clawdbot setups as headless coding servers. One developer: “starting Claude Code sessions on my phone while putting the baby to bed - to then commit and review later.” The setup gives Clawdbot “the ability to command Claude Code / any coding agent, so it has the ability to extend itself infinitely with any tool.”
Why you should care
This isn’t a tech demo. People are replacing Zapier automations with local cron jobs. They’re coding from their phones. One user’s take: “For a relatively small amount of $$ we can have an insanely capable always-on AI assistant. Feels like cheating.“
The productivity divide
$599 down + $100-200/month in API costs buys you a 24/7 AI assistant that codes, automates, and extends itself. Those who can afford the setup get (allegedly) massive leverage. Those who can’t are now competing against people who do.
Direct quotes from users:
“Feels like cheating.“
“At this point, if you don’t have @clawdbot and at least 3 long running agents, you are not hireable“
“It’s running my company.” [Really?]
Translation
AI advantage isn’t about using ChatGPT better. It’s about deploying always-on agents with filesystem access, self-modification capabilities, and the ability to spawn coding sessions on demand. The barrier to entry is a Mac Mini.
🧠 Claude’s New Constitution: Teaching Ethics Through Understanding
What happened
Anthropic released Claude’s new constitution under Creative Commons CC0. Instead of a list of rules, the constitution explains the “why” behind desired behaviors, enabling Claude to exercise judgment across novel situations.
Why it matters
Previous approaches gave AI systems rules: “Don’t do X.” The new approach explains reasoning: “Here’s why X causes harm, and here’s how to think about similar situations.” It’s the difference between teaching multiplication tables and teaching how numbers work. The constitution itself is primarily written for Claude to read and understand, not for human reference.
The approach
Four core priorities in order: broadly safe, broadly ethical, compliant with guidelines, genuinely helpful. The constitution is used throughout training to generate synthetic training data, with Claude itself constructing examples based on understanding the principles.
Why to be cautious
Anthropic acknowledges a gap between the constitution’s ideals and actual model behavior. Training toward the vision remains an ongoing technical challenge. Future powerful models may fail even if current training succeeds. Teaching “why” is harder than enforcing “what” - but potentially more robust if it works.
The lesson
AI safety isn’t solved by longer rule lists. It requires systems that understand intent, can generalize principles to new contexts, and exercise judgment. Whether that’s achievable through training on principles remains an open question - but it’s a more promising path than endlessly enumerating edge cases.
🛠️ The Skills Ecosystem: NPM for Agent Capabilities
Three pieces converged this week to create a standard way to extend AI agents:
Add-Skill: Install Agent Capabilities Like NPM Packages
Vercel Labs released add-skill, a CLI tool for installing agent skills from any git repository with a single command. Skills are reusable instruction sets defined in SKILL.md files with YAML frontmatter, compatible with 25+ coding agents including Claude Code, Cursor, OpenCode, and GitHub Copilot.
Install with: npx add-skill vercel-labs/agent-skills
Skills.sh: The Leaderboard for Agent Capabilities
Skills.sh launched as a directory and leaderboard tracking install counts across the ecosystem. Top skill: vercel-react-best-practices with 39.3K installs. Total tracked: 12,180 skills.
The agentskills.io Specification
All of this works because agents are converging on a standard: SKILL.md files with metadata in YAML frontmatter. Cross-agent compatibility means you install once, use everywhere.
Why you should care
This is the moment agent capabilities become composable. Need React best practices? Install a skill. Need Remotion video creation patterns? Install a skill. Need your company’s coding standards? Write a skill once, distribute it to the team. The ecosystem already has 12,180 skills and clear adoption leaders. This is how tribal knowledge becomes portable.
🔬 AI-Resistant Technical Evaluations: The Arms Race with Claude
What happened
Anthropic’s performance engineering team published lessons from redesigning their technical take-home test three times as each Claude release defeated the previous version. These are the tests that potential employees have to undertake as part of the recruitment process. If candidates can use Claude to generate correct answers, the tests suddenly have no value. Claude Opus 4.5 now matches even the strongest human candidates within time limits.
The evolution
Original: 4-hour performance optimization challenge
Claude Opus 4 matched most humans
Opus 4.5 matched top candidates
Three iterations: removed multicore, focused on compression, moved to unconventional instruction sets
New version: Zachtronics-style puzzles with unusual, heavily constrained instruction sets
Why you should care
When your hiring process needs to stay ahead of your own AI models, you’re in an arms race with yourself. The team’s conclusion: evaluations must become increasingly unusual and out-of-distribution. Realism gets sacrificed for AI-resistance. The new test “simulates novel work” rather than resembling the actual job. As they note, this may be “a luxury we no longer have”.
The pattern
Every field will face this. How do you assess human capability when AI can match the strongest candidates in constrained timeframes? The answer: you test for increasingly unusual scenarios that AI hasn’t seen. The cost: tests that look less like actual work. The implication: human value shifts toward handling novel, weird, out-of-distribution situations.
🎯 Rapid Hits
💻 NVIDIA PersonaPlex: Full-Duplex Voice AI with Roles
NVIDIA released PersonaPlex-7B-v1, a real-time speech-to-speech model that listens and speaks simultaneously. Supports natural interruptions, barge-ins, and rapid turn-taking. Conditioned on both voice prompts (how to speak) and text prompts (what role to play). Commercial-ready under NVIDIA Open Model License, supporting 8 languages.
Why it matters: Voice AI that can handle actual conversational dynamics - overlapping speech, interruptions, context switching - in real-time. Your next customer service bot sounds natural, not robotic.
📱 Liquid AI: On-Device Reasoning Under 1GB
Liquid AI released LFM2.5-1.2B-Thinking, a reasoning model optimised for edge deployment running under 1GB memory. 239 tok/s on AMD CPU, 82 tok/s on mobile NPU. Rivals much larger models through 28T token pre-training and large-scale reinforcement learning.
Why it matters: Capable reasoning AI that runs on your phone with no cloud connection. Privacy-preserving, low-latency, offline-capable. The 1.2B model outperforms many 7B+ models while fitting in your pocket.
🔧 Ollama Adds Anthropic API Compatibility
Ollama introduced compatibility with Anthropic’s Messages API, enabling tools like Claude Code to use local models. Quick setup: ollama launch claude prompts model selection and auto-configures. Run Claude Code with local qwen3-coder or cloud models.
Why it matters: Cloud-quality coding assistance with local models. Privacy for sensitive codebases, no API costs for volume work, full control over your development environment.
🎬 Remotion: Create Videos Programmatically with React
Remotion enables creating real MP4 videos with React code. Parametrise content, render server-side, scale with Remotion Lambda. Now includes Claude Code integration for AI-assisted video project generation.
Why it matters: Marketing videos, year-in-review animations, social media content - all generated programmatically from your data. No video editing skills required.
💾 AirLLM: Running 70B Models on 4GB GPUs
AirLLM enables running 70B models on single 4GB GPU cards through inference memory optimization. No quantization, distillation, or pruning required. 4-bit/8-bit compression options provide 3x speed improvement. Runs 405B Llama3.1 on 8GB VRAM.
Why it matters: Capable models on consumer hardware. No cloud dependencies, no API costs, full control. The barrier to running frontier-class models just dropped to a gaming GPU.
📖 Claude Code Best Practices: The Comprehensive Guide
Anthropic released comprehensive best practices for Claude Code covering configuration, communication, session management, and scaling. Core insight: context window is the most important resource to manage. Four-phase workflow: Explore (Plan Mode) → Plan → Implement (Normal Mode) → Commit.
Key recommendations: CLAUDE.md for persistent context, skills for domain knowledge, hooks for deterministic actions, /clear frequently between unrelated tasks, give Claude verification criteria (tests, screenshots, expected outputs).
Why it matters: The difference between Claude Code feeling magical and feeling frustrating is understanding context management. This guide is the operating manual.
🚨 The White House Started Manipulating Arrest Photos
What happened
The Guardian revealed the White House posted a digitally altered image of Nekima Levy Armstrong’s arrest, changing her neutral expression to appear crying and darkening her skin tone. The original unaltered image was posted by DHS Secretary Kristi Noem 30 minutes earlier.
The details
DHS posted original at 10:21am, White House posted altered version at 10:51am
Guardian overlay analysis confirmed images are identical except for facial alterations
No disclosure that image was manipulated
Deputy communications director’s response: “Enforcement of the law will continue. The memes will continue.”
The stakes
When the White House communications shop treats AI manipulation as routine and promises “the memes will continue”, the question isn’t whether it’s wrong. It’s whether anyone has the power to stop it.
🚀 Your Weekend Project
Pick one:
Set up Clawdbot - If you have a Mac (or Linux), install Clawdbot and connect it to Telegram. Start simple: have it check the weather, set a reminder, or send you a daily summary. Then try something more ambitious: ask it to automate a workflow you currently do manually.
Install your first agent skill - Run
npx add-skill vercel-labs/agent-skillsif you use Claude Code, Cursor, or a compatible agent. Browse skills.sh for capabilities relevant to your work. Watch how your agent’s behavior changes with domain knowledge loaded.Read Claude’s constitution - Open Anthropic’s new constitution and skim through it. Pay attention to the “why” explanations, not just the rules. Think about how you’d explain your own values to an AI - principles vs. prohibitions.
Try Ollama + Claude Code locally - If you have decent hardware, install Ollama and run
ollama launch claude. Point it at qwen3-coder and run a coding task entirely locally. Feel the difference between cloud-dependent and self-hosted.Personal AI - Ask yourself: what would change if I had a 24/7 AI assistant with filesystem access?
🏗️ About Barnacle Labs
At Barnacle Labs we build AI systems that actually ship. From the National Cancer Institute’s NanCI app to AI systems deployed across biotech and enterprise clients, we’re the ‘breakthroughs, not buzzwords’ team.
Got an AI challenge that’s stuck? Reply to this email—let’s talk.
The voices worth listening to in AI are the ones building, not just talking. See you next week.