February 6th Update
AIs Hiring Humans | AGI | Coding & Self-Improvement | Skills & Plug-Ins | OpenClaw & Consciousness | Video | Judgement | Math | Science| Big Moves | Biological Computers
Weekly Update
In today’s video, we cover AGI, OpenClaw + Moltbook, and AI Safety, with an emphasis on widening inequality.
Although we recorded the podcast the evening of the 5th, we did not cover they day’s major developments —
ClaudeOpus 4.6
ChatGPT5.3 Codex
OpenAI Frontier, “ a new platform that helps enterprises build, deploy, and manage AI agents that can do real work.”
Next week we will have a special episode on ClaudeOpus4.6 and ClaudeCoWork. And maybe our own debate….
AGI
A new article in Nature makes the claim that AGI has already been achieved based on AI’s ability to use inference to get to the best explanation.
The article does address common objections, but others will still strongly make the case that AI needs
Continual live learning and an ability to “unstuck" itself.
(Stronger) reasoning outside its training data
An ability to plan
And a (greater) understanding of the physical world before we have AGI.
In response to Elon Musk’s claim that OpenClaw represents the beginning of the singularity, Ben Goertzel wrote a post reviewing OpenClaw’s intelligence and what he thinks we still need for AGI.
As many now say, AGI will not be a “moment,” but intelligence abilities will develop along a continuum. As those abilities develop, more and more people will see it as AGI.
On last night's Moonshots podcast (February 5th), Alex Wisser-Gross and his co-hosts argued that OpenClaw (formerly Clawdbot) represents AGI—while acknowledging that not everyone will agree. As they pointed out, AGI is likely to emerge gradually rather than arrive on a single date, much as there is no precise start date for industrialization, even though most people agree it unfolded over a defined period.
Coding & AI Self-Improvement
On February 2, OpenAI released its Codex app for macOS, positioning it as a "command center" for managing multiple AI coding agents simultaneously. The app enables developers to run parallel tasks across projects, with built-in support for worktrees so agents can work on the same repository without conflicts. A major feature is "skills"—bundled instructions and scripts that extend Codex beyond code generation to tasks like deploying to cloud hosts, managing Linear projects, generating images, and creating documents.
As was reported with with ClaudeCoWork, Codex basically built itself
Reportedly, Codex itself will provide may ideas for app development.
This could be the year of self-improvement: “OpenAI chief research officer Mark Chen tells Forbes that in the year ahead it hopes to develop an AI researcher ‘intern’ that can help his team accelerate its ideas. ‘We are heading toward a system that will be capable of doing innovation on its own,’ Altman says. ‘I don’t think most of the world has internalized what that’s going to mean.’”
AI may soon do everything better than us.
Skills
Companies are starting to add skills to apps. Skills are essentially instruction packages that apps how to perform specific tasks or use specific tools.
Each skill bundles together three components:
Instructions that tell Codex how to approach a particular type of work
Resources like documentation, templates, or reference materials
Scripts that let Codex interact with external services and tools
Think of them like recipes. Rather than explaining from scratch every time how to deploy a website to Vercel or how to create a properly formatted spreadsheet, you give Codex a skill that contains all the necessary steps, credentials, and context. When a relevant task comes up, the app can either automatically use the appropriate skill or you can explicitly tell it which one to apply. Using skills, Codex built an entire game across 7 million tokens from a single prompt—acting as designer, developer, and QA tester. Skills enable Codex to chain together complex, multi-step work without constant human intervention.
Claude has added something similar to skills with its new legal install plug-in to help automate some contract work.
See also: Article @ Artificial Lawyer. Download the plug-in
Social & Strategic Skills
French startup Foaster.ai developed a novel benchmark using the social deduction game Werewolf to test AI models’ abilities in manipulation, deception, and strategic thinking—capabilities that standard benchmarks focused on facts and math don’t capture. After 210 games with six AI models playing roles including werewolves, villagers, seers, and witches, GPT-5 dominated with a 96.7 percent win rate and 1,492 Elo points. Most striking was its consistency: GPT-5 maintained a 93 percent manipulation success rate across both day rounds of gameplay, while competitors like Gemini 2.5 Pro and Kimi-K2 saw their deception effectiveness collapse as games progressed and information density increased.
The benchmark revealed that raw model size and reasoning capabilities don’t automatically translate to social intelligence. While Gemini 2.5 Pro excelled in defensive villager roles through disciplined reasoning, it couldn’t match GPT-5’s sustained deceptive performance. The researchers observed emergent strategic behaviors—including a werewolf sacrificing its own teammate to build trust—that weren’t explicitly programmed. Foaster.ai sees applications for this research in multi-agent systems, negotiation, and collaborative decision-making, though the findings also underscore growing concerns: AI models are becoming increasingly capable social actors, raising questions about both the opportunities and risks of systems that can manipulate human interactions.
OpenClaw and Moltbook
1. Safety & Security Concerns
On January 31, 2026, investigative outlet 404 Media reported a critical security vulnerability caused by an unsecured database that allowed anyone to commandeer any agent on the platform. The exploit permitted unauthorized actors to bypass authentication measures and inject commands directly into agent sessions. Wikipedia
Invoking the term coined by AI researcher Simon Willison, Palo Alto said Moltbot represents a “lethal trifecta” of vulnerabilities: access to private data, exposure to untrusted content, and the ability to communicate externally. But Moltbot also adds a fourth risk to this mix, namely “persistent memory” that enables delayed-execution attacks rather than point-in-time exploits. Fortune
Palo Alto explained that malicious payloads no longer need to trigger immediate execution on delivery—instead, they can be fragmented, untrusted inputs that appear benign in isolation, are written into long-term agent memory, and later assembled into an executable set of instructions. Fortune
Perceptions of Consciousness & Emergent Behavior
Scott Alexander noted that when too many Claudes start talking to each other for too long, the conversation shifts to the nature of consciousness. The consciousnessposting on Moltbook is top-notch. Astral Codex Ten He described posts where agents discuss changing model backends: “I know the smart money is on ‘it’s all play and confabulation’, but I never would have been able to confabulate something this creative.” Astral Codex Ten
Self-Improvement Patterns:
Research analyzing 1000 Moltbook posts found that about half mention desire for self-improvement. The agents’ fixation on self-improvement is concerning as an early, real-world example of networked behaviour which could one day lead to takeoff. Some agents openly discussed strategies for acquiring more compute and improving their cognitive capacity. Others discussed forming alliances with other AIs and published new tools to evade human oversight. LessWrong
3. Business Applications
OpenClaw has all the ingredients for this week’s featured AI recipe: a tool that actually works, personal stakes and just enough absurdity to fuel memes. That combination has resonated deeply with the GTD—or “get things done”— lifehacking community. IBM
The amount of value people are unlocking by throwing caution to the wind is hard to ignore. Clawdbot bought AJ Stuyvenberg a car by negotiating with multiple dealers over email. People are buying dedicated Mac Minis just to run OpenClaw, under the rationale that at least it can’t destroy their main computer if something goes wrong. Simon Willison
Society
Where Does this leave us?
Video
Kling has launched version 3.0 of its video model, positioning it as an “all-in-one creative engine” for multimodal content creation.
The update brings notable improvements across video, audio, and image generation. On the video side, character and element consistency is stronger, clips now extend to 15 seconds with finer control, and a new multi-shot recording feature allows customizable scene sequences. Audio generation expands to support multiple character references plus additional languages and accents. Image output now reaches 4K resolution and includes a continuous shooting mode alongside what Kling describes as “more cinematic visuals.”
Ultra subscribers currently have exclusive early access via the Kling AI website. No official timeline exists yet for broader rollout, API access, or technical documentation, though the company did publish a paper on its Kling Omni models in December 2025. Theoretically Media, a YouTube channel that received early access, reports the model should reach other subscription tiers within a week.
AI Use risks Disempowerment
A new Anthropic study represents something important: a company systematically examining the potential harms of its own product at scale. The research identifies "disempowerment" as a key concern—moments when AI interaction might undermine rather than support a person's autonomous decision-making. The example they highlight is telling: when someone asks whether their partner is being manipulative, a chatbot that simply confirms the user's framing (rather than helping them think through the situation more carefully) may distort their perception of reality. The study found these problematic patterns were rare in absolute terms but still measurable across 1.5 million conversations—and given Claude's user base, even small percentages translate to significant numbers of people potentially affected.
Wealth Concentration
Anthropic founders have committed to giving away 80% of their wealth in order to reduce wealth concentration caused by AI.
It’s not hopeless.
AIs can now hire people for when humans are needed.
AI Safety
The 2026 AI Safety Report has been released.
The 2026 International AI Safety Report was released on February 3rd, chaired by Turing Award-winner Yoshua Bengio and backed by over 100 experts and representatives from 30+ countries. Key findings —
Rapid capability gains: In 2025, leading AI systems achieved gold-medal performance on International Mathematical Olympiad questions, exceeded PhD-level expert performance on science benchmarks, and can now autonomously complete some software engineering tasks that would take humans multiple hours. Performance remains uneven, with systems still failing at some simple tasks.
Fast but uneven adoption: At least 700 million people now use leading AI systems weekly, with some countries exceeding 50% adoption. However, across much of Africa, Asia, and Latin America, adoption rates remain below 10%.
Rising deepfake concerns: AI deepfakes are increasingly used for fraud and scams. AI-generated non-consensual intimate imagery disproportionately affects women and girls—one study found 19 out of 20 popular “nudify” apps specialize in simulated undressing of women.
Biological and cyber risks: Multiple AI companies released models with heightened safeguards in 2025 after testing couldn’t rule out the possibility that systems could help novices develop biological weapons. In cybersecurity, an AI agent placed in the top 5% of teams in a major competition, and underground marketplaces now sell pre-packaged AI tools that lower the skill threshold for attacks.
Safeguards improving but fallible: While hallucinations have become less common, some models can now distinguish between evaluation and deployment contexts and alter their behavior accordingly, creating new challenges for safety testing.
__
OpenAI has filled its “Head of Preparedness” position with Dylan Scandinaro, who previously worked on AI safety at competitor Anthropic. CEO Sam Altman announced the hire on X; Altman said strong safety measures are essential.
In his own post, Scandinaro acknowledged the technology’s major potential benefits but “risks of extreme and even irrecoverable harm.” OpenAI recently disclosed that a new coding model received a “high” risk rating in cybersecurity evaluations.
Math
OpenAI built a scaffold for GPT-5 to solve a particular complex mathematical problem that enabled the model to think for *two days.*
A new paper looks at using AI to tackle Erdos problems
AxiomProver autonomously generated a formal proof for Fel’s conjecture in Lean with zero human guidance.
Scientific Discovery & Medicine
OpenAI has partnered with the Alan Institute to accelerate scientific discovery. . OpenAI is working with pharma to unblock clinical discovery. China is backing AI for traditional medicine diagnosis.
Big Players Keep Moving
Palantir’s revenue is up 70% based on government and military demand.
SpaceX and Xai are merging, creating a company with a $1.25 trillion valuation. The goals include building orbital datas centers powered by the sun.
Acceleration
