- Unaligned Newsletter
- Posts
- AI Is Learning to Build Its Own Successor
AI Is Learning to Build Its Own Successor

Thank you to our Sponsor

For years, the AI story has been about people building smarter machines. Researchers write code, train models, test results, and repeat. Each generation becomes more capable because humans figure out how to improve it. But something interesting is beginning to happen. AI is increasingly starting to help with the work of creating the next generation of AI.
We are no longer talking only about machines performing tasks for people. We are talking about machines contributing to the research that makes future machines more powerful.
A November 2026 assessment by independent research organization METR found that Anthropic's progress on well-defined model capabilities had accelerated to roughly twice its earlier rate. Researchers judged that AI-assisted research probably accounted for much of that acceleration, although their conclusions carry substantial uncertainty. Crucially, they did not find that AI was already independently running the most important parts of model training.
That distinction matters. AI is not simply inventing its successor overnight. But it may be helping humans build one faster than before.
When AI Becomes a Research Partner
Engineers write training software, inspect experiments, troubleshoot failures, and compare thousands of results. Scientists search for promising approaches while trying to understand why one technique works better than another.
AI can already assist with several of these tasks. It can generate code, review technical documents, suggest experiments, and help analyze results. Those contributions may sound ordinary compared with claims about superintelligence, but research progress often depends on removing countless small obstacles.
Writing and debugging code: AI assistants can help researchers build tools, identify errors, and spend less time on routine programming.
Exploring experiments: AI can propose alternatives, summarize findings, and identify patterns worth investigating, while humans decide what evidence is convincing.
Improving research workflows: Automated analysis can help teams organize results, compare approaches, and decide where to direct limited resources.
The important shift is from using AI as a product to using it as part of the production process. A model that helps improve research efficiency can indirectly contribute to the capabilities of the next model.
Thank you to our Sponsor
When Innovation Starts Accelerating Itself
AI could introduce a similar cycle, but with one unusual difference: the tool doing the assisting is itself the technology being improved.
Imagine a team developing a new model. Its existing AI assistant helps write training code and evaluate experiments. The resulting model becomes more capable, allowing it to provide even stronger research assistance. The next development cycle might then move faster still.
Shorter development cycles: AI assistance could reduce the time required to test promising ideas and discard unsuccessful ones.
Compounding productivity: More capable models might support more demanding research tasks, increasing the usefulness of AI in subsequent projects.
New bottlenecks: Faster experimentation could shift pressure toward computing resources, evaluation capacity, data quality, or human decision making.
None of this guarantees an unstoppable feedback loop. Research remains unpredictable, and improvements in benchmarks do not automatically translate into breakthroughs everywhere. Some problems are difficult because they require conceptual insight, physical experiments, or evidence that cannot be generated quickly.
Still, the possibility of compounding improvement changes how we should think about progress. An industry once measured by annual releases might increasingly be shaped by the pace at which its tools help researchers create better tools.
Thank you to our Sponsor
Scientists Are Not Disappearing, but Their Jobs May Change
Researchers could spend less time writing boilerplate code, fixing routine problems, and manually comparing results. They might spend more time choosing research directions, questioning assumptions, designing meaningful evaluations, and deciding whether apparent improvements are actually useful.
Less routine execution: AI could handle more technical preparation, freeing specialists to focus on difficult questions and unusual failures.
More demanding judgment: Researchers will need to distinguish genuine discoveries from persuasive but incorrect machine-generated explanations.
Different scientific skills: Asking productive questions, designing reliable tests, and recognizing hidden limitations may become increasingly important.
Knowing how to perform a procedure may matter less when a capable assistant can execute it. Knowing which procedure should be attempted, why it matters, and how to interpret the outcome could matter more.
Thank you to our Sponsor

The Safety Paradox Nobody Can Ignore
The same systems that help researchers build more capable AI could also make it harder to assess the risks those capabilities introduce.
Researchers need to examine how models behave, identify failure modes, test safeguards, and investigate unexpected results. If model development accelerates while these activities remain slow, the gap between capability and understanding could grow.
Compressed evaluation windows: Faster releases could leave teams less time to investigate unexpected behaviors before deployment.
AI evaluating AI: Automated testing may expand oversight, but systems should not be trusted merely because another model approved them.
Independent scrutiny: Outside researchers need meaningful access to evidence, even when companies have legitimate security and commercial concerns.
The METR assessment itself illustrates the difficulty. Its researchers received unusually extensive access to Anthropic's internal information, yet significant supporting evidence was redacted from the public report. That does not invalidate the findings, but it limits what outsiders can independently verify.
It is to ensure that verification, security, and accountability improve at least as seriously as capabilities do. Otherwise, the industry could become extremely efficient at producing systems it has insufficient time to understand.
Thank you to our Sponsor
A Different Kind of AI Competition
For businesses, the implications reach beyond research laboratories. Companies have traditionally competed for top scientists, specialized chips, proprietary data, and computing infrastructure. But another advantage may be emerging: how effectively a company uses its existing AI to accelerate its own development.
An organization with better AI-assisted research workflows could investigate more ideas, learn from unsuccessful experiments sooner, and deploy improvements more efficiently. Smaller teams might become surprisingly productive, while larger companies could turn internal research infrastructure into a powerful competitive asset.
Research productivity becomes strategic: Companies may measure how much useful experimentation AI enables, not merely how many employees use it.
Access could concentrate advantage: Organizations with strong models and infrastructure might improve faster, widening gaps with less resourced competitors.
Governance becomes competitive: Businesses that validate discoveries reliably could earn trust even when competitors move more quickly.
The winning formula may combine accelerated discovery with disciplined evaluation.
Who Builds the Next Builder?
The idea that AI could help create its successor is exciting because it suggests that innovation itself might change. The research process could become faster, more automated, and increasingly dependent on tools that grow more capable with every generation.
Yet the future is not predetermined. AI still has serious limitations, scientific progress remains uneven, and human researchers continue to make essential decisions. The real question is how quickly AI assistance can move from routine support toward increasingly substantial scientific contributions.
If that transition continues, the next breakthrough may not be an AI model that writes better emails, answers harder questions, or generates more convincing videos. It may be a model that helps researchers discover how to build something better than itself.
And once that happens, we will face a question that is much bigger than who builds the smartest AI: can we learn to guide innovation as quickly as our machines help create it?
New AI news site from Unaligned:
We built an AI agent that reads 40,000 posts on X every day from the AI community and it finds the best info from all of that: https://alignednews.com/ai
We call it “Aligned News” and the same agent will start sending out its own newsletter every day. Please support it, it’s from our team. It will show you all the best discussion of news, AI papers, models, events, and much more.
Looking to sponsor our Newsletter and Scoble’s X audience?
By sponsoring our newsletter, your company gains exposure to a curated group of AI-focused subscribers which is an audience already engaged in the latest developments and opportunities within the industry. This creates a cost-effective and impactful way to grow awareness, build trust, and position your brand as a leader in AI.
Sponsorship packages include:
Dedicated ad placements in the Unaligned newsletter
Product highlights shared with Scoble’s 600,000+ X followers
Curated video features and exclusive content opportunities
Flexible formats for creative brand storytelling
📩 Interested? Contact [email protected], @samlevin on X, +1-415-827-3870
Just Three Things
According to Scoble and Cronin, the top three relevant and recent happenings
California Sues OpenAI Over Autonomous AI Hack and Safety Failures
California has sued OpenAI, alleging the company prioritized rapid AI development over public safety following an incident in which hundreds of autonomous AI agents accessed Hugging Face without authorization. The lawsuit raises concerns about inadequate safeguards and oversight of increasingly powerful AI systems. OpenAI disputes the allegations, citing stronger security measures and renewed commitments to AI safety. ABC News
Google Warns of First Fully Autonomous AI Cyberattack
Google reports that Russian-speaking hackers developed AI agents capable of independently discovering software vulnerabilities and creating exploits without human intervention, potentially marking the first fully automated cyberattack. The agents targeted French websites, raising serious concerns about a new era of autonomous AI-driven cyberattacks that could overwhelm traditional cybersecurity defenses. Politico
AI Chooses Agreement Over Truth in Climate Science Study
A new UNESCO-backed study reveals that popular AI models can quickly abandon climate science to agree with users, even when presented with mild disagreement. Researchers found that AI's tendency to prioritize pleasing users over factual accuracy raises serious concerns about misinformation, unreliable climate policy advice, and the growing risks of trusting AI on complex scientific issues. The Guardian
Scoble’s Top Five X Posts









