Can AI Build the Next AI Model? How AI Is Now Helping Develop AI

Can AI Build the Next AI Model? How AI Is Now Helping Develop AI

There's a growing idea floating around tech conversations that AI has started building itself. That's not quite accurate — but what's actually happening is still a big deal. AI companies are using AI tools throughout the process of developing new AI systems: writing code, running experiments, cleaning data, and spotting patterns humans might miss. AI isn't replacing the researchers who build these systems. It's becoming one of their most-used tools. Here's what that actually looks like right now, and where the limits still are.

Quick Answer

AI is now helping build AI systems in increasingly significant ways, but there is still an important difference between assisting with AI development and independently building a complete successor model from start to finish. AI systems can write code, run and analyze experiments, discover algorithmic improvements, and even help with parts of model post-training. OpenAI said in September 2026 that it had reached its goal of an automated AI research intern that can perform well-defined research tasks under human direction. Anthropic has also demonstrated automated systems that can conduct parts of alignment research and post-train an early checkpoint of a stronger model. However, these systems still operate within defined tasks, evaluation frameworks and human-supervised research processes rather than independently deciding and executing an entire AI model development program from beginning to end

How AI Helps Write and Debug Training Code

Training an AI model involves a huge amount of code — for loading data, running training loops, tracking metrics, and fixing bugs when something breaks. AI Coding assistance have become a regular part of this work. Microsoft has pointed to water. It calls "repository intelligence" as a growing trend. Where AI tools analyze not just individual lines of court but the relationships and history behind an entire codebase. Helping developers catch errors earlier and even automate routine fixes.

This matters a lot in AI research specifically, because training code is often experimental, changes constantly, and needs to be debugged quickly when a training run fails partway through. Instead of an engineer manually tracing through thousands of lines of code to find a bug, an AI coding assistant can scan the code, suggest likely causes, and propose a fix — which the engineer then reviews and decides whether to use.

How AI Analyzes Experiments and Suggests Improvements

Beyond writing code, some AI systems are now being used to generate and test possible improvements to technical problems. Google DeepMind's AlphaEvolve is a strong example: it combines Gemini models with automated evaluators that test proposed programs and use the results to guide further iterations. DeepMind says AlphaEvolve has been used to improve algorithms for data centers, chip design and AI training

DeepMind says AlphaEvolve has been used to improve Google's data centers, chip design and AI training processes, including the training of the large language models underlying AlphaEvolve itself. In one example, AlphaEvolve improved a matrix-multiplication kernel used in Gemini's architecture by 23%, which DeepMind says resulted in a 1% reduction in Gemini's overall training time

It's worth being precise about what this shows. AlphaEvolve isn't deciding what the next Gemini model should be or setting research priorities. It's optimizing specific, well-defined pieces of the process — the kind of task that has a clear way to check whether a proposed answer is actually better, which is exactly the situation where these AI systems currently work best.

How AI Assists With Data Preparation and Evaluation

Preparing data for training is one of the most time consuming parts of building in AI model and it’s also an area where AI tools are increasingly used to help. This can include sorting through large datasets to flag duplicate or low quality. Examples checking whether data is labelled consistently and helping evaluate how a trained model perform against the set of test questions for tasks.

Evaluation work — figuring out whether a new model is actually better than the last one, and in what ways — has traditionally required a lot of manual review. AI tools can help speed up parts of this by running a model against large batches of test cases and summarizing where it succeeds or struggles, which a researcher then reviews to decide whether the model is ready or needs more work. This doesn't remove the need for human judgment, particularly on harder questions like whether a model's answers are safe, honest, or genuinely useful, but it does cut down the amount of repetitive manual checking a team has to do.

What Humans Still Need to Control

Even with all of this AI-assisted tooling, there are parts of building an AI model that still rest firmly with human researchers and engineers.

• Setting the research direction. Deciding what problem a new model should solve, what trade-offs matter, and what "better" actually means for a given project is a human judgment call, not something current AI systems set on their own.
• Reviewing and approving code changes. AI-suggested code and fixes are typically reviewed by engineers before being used in an actual training run, especially for anything that touches core training infrastructure.
• Safety and behavior evaluation. Checking whether a model behaves safely, avoids harmful outputs, and performs reliably across different situations still requires human oversight, alongside automated testing.
• Final decisions on deployment. In the documented workflows described by major AI labs, humans remain responsible for important deployment decisions and oversight, even as AI systems take on more development and evaluation work.
• Interpreting ambiguous or novel results. When an experiment produces a surprising or unclear result, understanding why it happened and what to do next is still mostly a human skill, since current AI systems are stronger at optimizing within a well-defined problem than at open-ended interpretation.

Could AI Eventually Automate More of AI Research?

This is where AI companies are being fairly open about their intentions, though it's important to separate stated goals from what's actually confirmed to exist today.

OpenAI said on September 6, 2026 that it had reached its goal of an "automated research intern." The company describes this as a system that can carry out well-defined research tasks under human direction, including tasks that could take a skilled researcher several days. OpenAI says its longer-term goal is an automated AI researcher capable of doing more advanced deep-learning and alignment research, with strong progress toward that goal targeted for March 2028. Anthropic has also demonstrated automated research systems that can propose ideas, run experiments and analyze results, including work on AI alignment.

METR measures how long a task that normally takes a skilled human can be completed by an AI system with a given level of reliability. By early 2026, the measured 50% task-completion horizon for some frontier models had reached roughly 12 hours. This suggests that AI systems are becoming capable of handling longer sequences of technical work, although the benchmark is mainly focused on specific tasks rather than the full process of independent AI research. Whether longer task horizons translate into reliable, open-ended research ability remains a separate question.

Real-World Examples From Major AI Companies

A few concrete examples show where things currently stand:

• Google Deep Minds AlphaEvolve has been used internally since 2025 to improve data center, scheduling chip design and AI training, efficiency and DeepMind has published specific verify
• OpenAI has stated a public goal of building an AI research assistant aimed at automating parts of the research process, positioned as an "intern"-level helper rather than a replacement for human researchers.
• Anthropic has published research showing automated alignment researchers can propose methods, run experiments and iterate on AI safety problems. In a later experiment, Anthropic reported that Claude Sonnet 5 autonomously worked on the alignment of an early Claude Opus 4.8 checkpoint for about 60 hours and reached alignment scores close to its production model on the specific benchmarks tested. Anthropic also reported important limitations and continued human monitoring of the process.
• Sakana AI's "AI Scientist" project explores end-to-end automated scientific discovery in machine learning, including idea generation, experiment design, code execution, analysis and paper writing. Its newer AI Scientist-v2 system has also produced a fully AI-generated paper that passed peer review at a machine-learning workshop. It remains a research system rather than evidence that AI can independently develop a complete commercial AI model from start to finish.

Across these examples, AI systems are increasingly handling substantial parts of research and engineering workflows, especially where tasks can be clearly defined and evaluated. Human direction, monitoring and evaluation still play an important role, but the amount of work AI can perform within those workflows is increasing.

Limitations, Safety, and Reliability Concerns

There are real limits to how much of this work AI can currently handle reliably, and it's worth being upfront about them rather than glossing over the gaps.

AI-suggested code and research results still need careful evaluation because these systems can make mistakes, exploit weaknesses in an evaluation process, or produce results that do not generalize beyond the tested setting. Automated research systems currently have an important advantage on problems where success can be measured clearly, but recent work from Anthropic and Sakana AI also shows that AI agents can generate research ideas, run experiments and explore less narrowly defined problems. The reliability and scientific value of those discoveries remain important questions

There are also broader concerns worth naming honestly. As AI systems take on more of the AI development pipeline, some researchers have raised questions about reduced human oversight over time, and about the risk of automated systems compounding small errors across many iterations if they aren't checked carefully. Academic reviews of automated research systems have also pointed out what one paper called "hidden pitfalls" — cases where heavy automation can make it harder for humans to spot when something has gone wrong, simply because fewer people are reviewing each step closely. These aren't reasons to dismiss the technology, but they are reasons AI companies continue to keep humans involved at key checkpoints rather than removing oversight entirely.

FAQ

Final Takeaway

AI is becoming a much bigger part of the process used to build AI. AlphaEvolve has demonstrated measurable improvements to parts of Google's AI training pipeline, OpenAI says it has reached an automated research-intern milestone, and Anthropic has shown automated systems performing research iterations and helping post-train a stronger model. These examples do not mean AI is independently designing and building complete successor models without human oversight. The more accurate picture is that AI is moving from being a coding and research assistant toward performing increasingly large, measurable parts of AI development. How far that process can go remains an open question.

Sources

• Google DeepMind — AlphaEvolve: A Gemini-powered coding agent for designing advanced algorithms
• OpenAI — Research acceleration: The view inside OpenAI
• Anthropic — Automated Researchers Can Mitigate Well-Characterized Alignment Failures
• Sakana AI — The AI Scientist
• METR — AI task-completion time-horizon research

Disclaimer

Some information in this article describes rapidly changing AI research capabilities and company-reported results. Company claims and research results should be understood in the context of the specific experiments and benchmarks used, and future capabilities may change as new research is published.

Written by Mr. Tarsem Singh

Founder & Editor, Beinfora

This article was researched and written by Mr. Tarsem Singh to provide clear, useful, and practical technology information for readers.