While the agent didn’t face an actual CAPTCHA puzzle with images in this case, successfully passing Cloudflare’s behavioral screening that determines whether to present such challenges demonstrates sophisticated browser automation.
To understand the significance of this capability, it’s important to know that CAPTCHA systems have served as a security measure on the web for decades. Computer researchers invented the technique in the 1990s to screen bots from entering information into websites, originally using images with letters and numbers written in wiggly fonts, often obscured with lines or noise to foil computer vision algorithms. The assumption is that the task will be easy for humans but difficult for machines.
Cloudflare’s screening system, called Turnstile, often precedes actual CAPTCHA challenges and represents one of the most widely deployed bot-detection methods today. The checkbox analyzes multiple signals, including mouse movements, click timing, browser fingerprints, IP reputation, and JavaScript execution patterns to determine if the user exhibits human-like behavior. If these checks pass, users proceed without seeing a CAPTCHA puzzle. If the system detects suspicious patterns, it escalates to visual challenges.
The ability for an AI model to defeat a CAPTCHA isn’t entirely new (although having one narrate the process feels fairly novel). AI tools have been able to defeat certain CAPTCHAs for a while, which has led to an arms race between those that create them and those that defeat them. OpenAI’s Operator, an experimental web-browsing AI agent launched in January, faced difficulty clicking through some CAPTCHAs (and was also trained to stop and ask a human to complete them), but the latest ChatGPT Agent tool has seen a much wider release.
It’s tempting to say that the ability of AI agents to pass these tests puts the future effectiveness of CAPTCHAs into question, but for as long as there have been CAPTCHAs, there have been bots that could later defeat them. As a result, recent CAPTCHAs have become more of a way to slow down bot attacks or make them more expensive rather than a way to defeat them entirely. Some malefactors even hire out farms of humans to defeat them in bulk.
For those who missed the Flash era, these in-browser apps feel somewhat like the vintage apps that defined a generation of Internet culture from the late 1990s through the 2000s when it first became possible to create complex in-browser experiences. Adobe Flash (originally Macromedia Flash) began as animation software for designers but quickly became the backbone of interactive web content when it gained its own programming language, ActionScript, in 2000.
But unlike Flash games, where hosting costs fell on portal operators, Anthropic has crafted a system where users pay for their own fun through their existing Claude subscriptions. “When someone uses your Claude-powered app, they authenticate with their existing Claude account,” Anthropic explained in its announcement. “Their API usage counts against their subscription, not yours. You pay nothing for their usage.”
A view of the Anthropic Artifacts gallery in the “Play a Game” section. Benj Edwards / Anthropic
Like the Flash games of yesteryear, any Claude-powered apps you build run in the browser and can be shared with anyone who has a Claude account. They’re interactive experiences shared with a simple link, no installation required, created by other people for the sake of creating, except now they’re powered by JavaScript instead of ActionScript.
While you can share these apps with others individually, right now Anthropic’s Artifact gallery only shows examples made by Anthropic and your own personal Artifacts. (If Anthropic expanded it into the future, it might end up feeling a bit like Scratch meets Newgrounds, but with AI doing the coding.) Ultimately, humans are still behind the wheel, describing what kinds of apps they want the AI model to build and guiding the process when it inevitably makes mistakes.
Speaking of mistakes, don’t expect perfect results at first. Usually, building an app with Claude is an interactive experience that requires some guidance to achieve your desired results. But with a little patience and a lot of tokens, you’ll be vibe coding in no time.
Anthropic says Claude 4 beats Gemini on coding benchmarks; works autonomously for hours.
The Claude 4 logo, created by Anthropic. Credit: Anthropic
On Thursday, Anthropic released Claude Opus 4 and Claude Sonnet 4, marking the company’s return to larger model releases after primarily focusing on mid-range Sonnet variants since June of last year. The new models represent what the company calls its most capable coding models yet, with Opus 4 designed for complex, long-running tasks that can operate autonomously for hours.
Alex Albert, Anthropic’s head of Claude Relations, told Ars Technica that the company chose to revive the Opus line because of growing demand for agentic AI applications. “Across all the companies out there that are building things, there’s a really large wave of these agentic applications springing up, and a very high demand and premium being placed on intelligence,” Albert said. “I think Opus is going to fit that groove perfectly.”
Before we go further, a brief refresher on Claude’s three AI model “size” names (first introduced in March 2024) is probably warranted. Haiku, Sonnet, and Opus offer a tradeoff between price (in the API), speed, and capability.
Haiku models are the smallest, least expensive to run, and least capable in terms of what you might call “context depth” (considering conceptual relationships in the prompt) and encoded knowledge. Owing to the small size in parameter count, Haiku models retain fewer concrete facts and thus tend to confabulate more frequently (plausibly answering questions based on lack of data) than larger models, but they are much faster at basic tasks than larger models. Sonnet is traditionally a mid-range model that hits a balance between cost and capability, and Opus models have always been the largest and slowest to run. However, Opus models process context more deeply and are hypothetically better suited for running deep logical tasks.
A screenshot of the Claude web interface with Opus 4 and Sonnet 4 options shown. Credit: Anthropic
There is no Claude 4 Haiku just yet, but the new Sonnet and Opus models can reportedly handle tasks that previous versions could not. In our interview with Albert, he described testing scenarios where Opus 4 worked coherently for up to 24 hours on tasks like playing Pokémon while coding refactoring tasks in Claude Code ran for seven hours without interruption. Earlier Claude models typically lasted only one to two hours before losing coherence, Albert said, meaning that the models could only produce useful self-referencing outputs for that long before beginning to output too many errors.
In particular, that marathon refactoring claim reportedly comes from Rakuten, a Japanese tech services conglomerate that “validated [Claude’s] capabilities with a demanding open-source refactor running independently for 7 hours with sustained performance,” Anthropic said in a news release.
Whether you’d want to leave an AI model unsupervised for that long is another question entirely because even the most capable AI models can introduce subtle bugs, go down unproductive rabbit holes, or make choices that seem logical to the model but miss important context that a human developer would catch. While many people now use Claude for easy-going vibe coding, as we covered in March, the human-powered (and ironically-named) “vibe debugging” that often results from long AI coding sessions is also a very real thing. More on that below.
To shore up some of those shortcomings, Anthropic built memory capabilities into both new Claude 4 models, allowing them to maintain external files for storing key information across long sessions. When developers provide access to local files, the models can create and update “memory files” to track progress and things they deem important over time. Albert compared this to how humans take notes during extended work sessions.
Extended thinking meets tool use
Both Claude 4 models introduce what Anthropic calls “extended thinking with tool use,” a new beta feature allowing the models to alternate between simulated reasoning and using external tools like web search, similar to what OpenAI’s o3 and 04-mini-high AI models currently do in ChatGPT. While Claude 3.7 Sonnet already had strong tool use capabilities, the new models can now interleave simulated reasoning and tool calling in a single response.
“So now we can actually think, call a tool process, the results, think some more, call another tool, and repeat until it gets to a final answer,” Albert explained to Ars. The models self-determine when they have reached a useful conclusion, a capability picked up through training rather than governed by explicit human programming.
General Claude 4 benchmark results, provided by Anthropic. Credit: Anthropic
In practice, we’ve anecdotally found parallel tool use capability very useful in AI assistants like OpenAI o3, since they don’t have to rely on what is trained in their neural network to provide accurate answers. Instead, these more agentic models can iteratively search the web, parse the results, analyze images, and spin up coding tasks for analysis in ways that can avoid falling into a confabulation trap by relying solely on pure LLM outputs.
“The world’s best coding model”
Anthropic says Opus 4 leads industry benchmarks for coding tasks, achieving 72.5 percent on SWE-bench and 43.2 percent on Terminal-bench, calling it “the world’s best coding model.” According to Anthropic, companies using early versions report improvements. Cursor described it as “state-of-the-art for coding and a leap forward in complex codebase understanding,” while Replit noted “improved precision and dramatic advancements for complex changes across multiple files.”
In fact, GitHub announced it will use Sonnet 4 as the base model for its new coding agent in GitHub Copilot, citing the model’s performance in “agentic scenarios” in Anthropic’s news release. Sonnet 4 scored 72.7 percent on SWE-bench while maintaining faster response times than Opus 4. The fact that GitHub is betting on Claude rather than a model from its parent company Microsoft (which has close ties to OpenAI) suggests Anthropic has built something genuinely competitive.
Software engineering benchmark results, provided by Anthropic. Credit: Anthropic
Anthropic says it has addressed a persistent issue with Claude 3.7 Sonnet in which users complained that the model would take unauthorized actions or provide excessive output. Albert said the company reduced this “reward hacking behavior” by approximately 80 percent in the new models through training adjustments. An 80 percent reduction in unwanted behavior sounds impressive, but that also suggests that 20 percent of the problem behavior remains—a big concern when we’re talking about AI models that might be performing autonomous tasks for hours.
When we asked about code accuracy, Albert said that human code review is still an important part of shipping any production code. “There’s a human parallel, right? So this is just a problem we’ve had to deal with throughout the whole nature of software engineering. And this is why the code review process exists, so that you can catch these things. We don’t anticipate that going away with models either,” Albert said. “If anything, the human review will become more important, and more of your job as developer will be in this review than it will be in the generation part.”
Pricing and availability
Both Claude 4 models maintain the same pricing structure as their predecessors: Opus 4 costs $15 per million tokens for input and $75 per million for output, while Sonnet 4 remains at $3 and $15. The models offer two response modes: traditional LLM and simulated reasoning (“extended thinking”) for complex problems. Given that some Claude Code sessions can apparently run for hours, those per-token costs will likely add up very quickly for users who let the models run wild.
Anthropic made both models available through its API, Amazon Bedrock, and Google Cloud Vertex AI. Sonnet 4 remains accessible to free users, while Opus 4 requires a paid subscription.
The Claude 4 models also debut Claude Code (first introduced in February) as a generally available product after months of preview testing. Anthropic says the coding environment now integrates with VS Code and JetBrains IDEs, showing proposed edits directly in files. A new SDK allows developers to build custom agents using the same framework.
A screenshot of “Claude Plays Pokemon,” a custom application where Claude 4 attempts to beat the classic Game Boy game. Credit: Anthropic
Even with Anthropic’s future riding on the capability of these new models, when we asked about how they guide Claude’s behavior by fine-tuning, Albert acknowledged that the inherent unpredictability of these systems presents ongoing challenges for both them and developers. “In the realm and the world of software for the past 40, 50 years, we’ve been running on deterministic systems, and now all of a sudden, it’s non-deterministic, and that changes how we build,” he said.
“I empathize with a lot of people out there trying to use our APIs and language models generally because they have to almost shift their perspective on what it means for reliability, what it means for powering a core of your application in a non-deterministic way,” Albert added. “These are general oddities that have kind of just been flipped, and it definitely makes things more difficult, but I think it opens up a lot of possibilities as well.”
Benj Edwards is Ars Technica’s Senior AI Reporter and founder of the site’s dedicated AI beat in 2022. He’s also a tech historian with almost two decades of experience. In his free time, he writes and records music, collects vintage computers, and enjoys nature. He lives in Raleigh, NC.
We’ve been expecting it for a while, and now it’s here: OpenAI has introduced an agentic coding tool called Codex in research preview. The tool is meant to allow experienced developers to delegate rote and relatively simple programming tasks to an AI agent that will generate production-ready code and show its work along the way.
Codex is a unique interface (not to be confused with the Codex CLI tool introduced by OpenAI last month) that can be reached from the side bar in the ChatGPT web app. Users enter a prompt and then click either “code” to have it begin producing code, or “ask” to have it answer questions and advise.
Whenever it’s given a task, that task is performed in a distinct container that is preloaded with the user’s codebase and is meant to accurately reflect their development environment.
To make Codex more effective, developers can include an “AGENTS.md” file in the repo with custom instructions, for example to contextualize and explain the code base or to communicate standardizations and style practices for the project—kind of a README.md but for AI agents rather than humans.
Codex is built on codex-1, a fine-tuned variation of OpenAI’s o3 reasoning model that was trained using reinforcement learning on a wide range of coding tasks to analyze and generate code, and to iterate through tests along the way.
It has recently been expanding its offerings—for example, it recently launched a deep research tool competing with similar ones provided by OpenAI and Google, as well as Sonar, an API for generative AI-powered search.
It will face fierce competition in the browser market, though. Google’s Chrome accounts for the majority of web browser use around the world, and despite its position at the forefront of AI search, Perplexity isn’t the first to introduce a browser with heavy use of generative AI features. For example, The Browser Company showed off its Dia browser in December.
Dia will allow users to type natural language commands into the search bar, like finding a document or webpage or creating a calendar event. It’s possible that Comet will do similar things, but again, we don’t know.
So far, most consumer-facing AI tools have come in one of three forms. There are general-purpose chatbots (like OpenAI’s ChatGPT and Anthropic’s Claude); features that use trained deep learning models subtly baked into existing software (as in Adobe Photoshop or Apple’s iOS); and, less commonly, standalone software meant to remake existing application categories using AI features (like the Cursor IDE).
There haven’t been a ton of AI-specific applications in existing categories like this before, but expect to see more coming over the next couple of years.
The researchers’ explanations about how “Set-of-Mark” and “Trace-of-Mark” work. Credit: Microsoft Research
The Magma model introduces two technical components: Set-of-Mark, which identifies objects that can be manipulated in an environment by assigning numeric labels to interactive elements, such as clickable buttons in a UI or graspable objects in a robotic workspace, and Trace-of-Mark, which learns movement patterns from video data. Microsoft says those features allow the model to complete tasks like navigating user interfaces or directing robotic arms to grasp objects.
Microsoft Magma researcher Jianwei Yang wrote in a Hacker News comment that the name “Magma” stands for “M(ultimodal) Ag(entic) M(odel) at Microsoft (Rese)A(rch),” after some people noted that “Magma” already belongs to an existing matrix algebra library, which could create some confusion in technical discussions.
Reported improvements over previous models
In its Magma write-up, Microsoft claims Magma-8B performs competitively across benchmarks, showing strong results in UI navigation and robot manipulation tasks.
For example, it scored 80.0 on the VQAv2 visual question-answering benchmark—higher than GPT-4V’s 77.2 but lower than LLaVA-Next’s 81.8. Its POPE score of 87.4 leads all models in the comparison. In robot manipulation, Magma reportedly outperforms OpenVLA, an open source vision-language-action model, in multiple robot manipulation tasks.
Magma’s agentic benchmarks, as reported by the researchers. Credit: Microsoft Research
As always, we take AI benchmarks with a grain of salt since many have not been scientifically validated as being able to measure useful properties of AI models. External verification of Microsoft’s benchmark results will become possible once other researchers can access the public code release.
Like all AI models, Magma is not perfect. It still faces technical limitations in complex step-by-step decision-making that requires multiple steps over time, according to Microsoft’s documentation. The company says it continues to work on improving these capabilities through ongoing research.
Yang says Microsoft will release Magma’s training and inference code on GitHub next week, allowing external researchers to build on the work. If Magma delivers on its promise, it could push Microsoft’s AI assistants beyond limited text interactions, enabling them to operate software autonomously and execute real-world tasks through robotics.
Magma is also a sign of how quickly the culture around AI can change. Just a few years ago, this kind of agentic talk scared many people who feared it might lead to AI taking over the world. While some people still fear that outcome, in 2025, AI agents are a common topic of mainstream AI research that regularly takes place without triggering calls to pause all of AI development.
On Tuesday, Hugging Face researchers released an open source AI research agent called “Open Deep Research,” created by an in-house team as a challenge 24 hours after the launch of OpenAI’s Deep Research feature, which can autonomously browse the web and create research reports. The project seeks to match Deep Research’s performance while making the technology freely available to developers.
“While powerful LLMs are now freely available in open-source, OpenAI didn’t disclose much about the agentic framework underlying Deep Research,” writes Hugging Face on its announcement page. “So we decided to embark on a 24-hour mission to reproduce their results and open-source the needed framework along the way!”
Similar to both OpenAI’s Deep Research and Google’s implementation of its own “Deep Research” using Gemini (first introduced in December—before OpenAI), Hugging Face’s solution adds an “agent” framework to an existing AI model to allow it to perform multi-step tasks, such as collecting information and building the report as it goes along that it presents to the user at the end.
The open source clone is already racking up comparable benchmark results. After only a day’s work, Hugging Face’s Open Deep Research has reached 55.15 percent accuracy on the General AI Assistants (GAIA) benchmark, which tests an AI model’s ability to gather and synthesize information from multiple sources. OpenAI’s Deep Research scored 67.36 percent accuracy on the same benchmark.
As Hugging Face points out in its post, GAIA includes complex multi-step questions such as this one:
Which of the fruits shown in the 2008 painting “Embroidery from Uzbekistan” were served as part of the October 1949 breakfast menu for the ocean liner that was later used as a floating prop for the film “The Last Voyage”? Give the items as a comma-separated list, ordering them in clockwise order based on their arrangement in the painting starting from the 12 o’clock position. Use the plural form of each fruit.
To correctly answer that type of question, the AI agent must seek out multiple disparate sources and assemble them into a coherent answer. Many of the questions in GAIA represent no easy task, even for a human, so they test agentic AI’s mettle quite well.
OpenAI is making ChatGPT work a little more like older digital assistants with a new feature called Tasks, as reported by TechCrunch and others.
Currently in beta, Tasks allows users to direct the chatbot to send reminders or to generate responses to specific prompts at certain times; recurring tasks are also supported.
The feature is available to Plus, Team, and Pro subscribers starting today, while free users don’t have access.
To create a task, users need to select “4o with scheduled tasks” from the model picker and then direct ChatGPT using the same kind of plain language text prompts that drive everything else it does. ChatGPT will sometimes suggest tasks, too, but they won’t go into effect unless the user approves them.
The user can then make changes to assigned tasks through the same chat conversation, or they can use a new Tasks section of the ChatGPT apps to manage all currently assigned items. There’s currently a 10-task limit.
When the time comes to perform an assigned task, the ChatGPT mobile or desktop app will send a notification on schedule.
This update can be seen as OpenAI’s first step into the agentic AI space, where applications built using deep learning can operate relatively independently within certain boundaries, either replacing or easing the day-to-day responsibilities of information workers.
Microsoft CEO Satya Nadella has announced a dramatic restructuring of the company’s engineering organization, which is pivoting the company’s focus to developing the tools that will underpin agentic AI.
Dubbed “CoreAI – Platform and Tools,” the new division rolls the existing AI platform team and the previous developer division (responsible for everything from .NET to Visual Studio) along with some other teams into one big group.
As for what this group will be doing specifically, it’s basically everything that’s mission-critical to Microsoft in 2025, as Nadella tells it:
This new division will bring together Dev Div, AI Platform, and some key teams from the Office of the CTO (AI Supercomputer, AI Agentic Runtimes, and Engineering Thrive), with the mission to build the end-to-end Copilot & AI stack for both our first-party and third-party customers to build and run AI apps and agents. This group will also build out GitHub Copilot, thus having a tight feedback loop between the leading AI-first product and the AI platform to motivate the stack and its roadmap.
To accomplish all that, “Jay Parikh will lead this group as EVP.” Parikh was hired by Microsoft in October; he previously worked as the VP and global head of engineering at Meta.
The fact that the blog post doesn’t say anything about .NET or Visual Studio, instead emphasizing GitHub Copilot and anything and everything related to agentic AI, says a lot about how Nadella sees Microsoft’s future priorities.
So-called AI agents are applications that are given specified boundaries (action spaces) and a large memory capacity to independently do subsets of the kinds of work that human office workers do today. Some company leaders and AI commentators believe these agents will outright replace jobs, while others are more conservative, suggesting they’ll simply be powerful tools to streamline the jobs people already have.
Over the past month, we’ve seen a rapid cadence of notable AI-related announcements and releases from both Google and OpenAI, and it’s been making the AI community’s head spin. It has also poured fuel on the fire of the OpenAI-Google rivalry, an accelerating game of one-upmanship taking place unusually close to the Christmas holiday.
“How are people surviving with the firehose of AI updates that are coming out,” wrote one user on X last Friday, which is still a hotbed of AI-related conversation. “in the last <24 hours we got gemini flash 2.0 and chatGPT with screenshare, deep research, pika 2, sora, chatGPT projects, anthropic clio, wtf it never ends."
Rumors travel quickly in the AI world, and people in the AI industry had been expecting OpenAI to ship some major products in December. Once OpenAI announced “12 days of OpenAI” earlier this month, Google jumped into gear and seemingly decided to try to one-up its rival on several counts. So far, the strategy appears to be working, but it’s coming at the cost of the rest of the world being able to absorb the implications of the new releases.
“12 Days of OpenAI has turned into like 50 new @GoogleAI releases,” wrote another X user on Monday. “This past week, OpenAI & Google have been releasing at the speed of a new born startup,” wrote a third X user on Tuesday. “Even their own users can’t keep up. Crazy time we’re living in.”
“Somebody told Google that they could just do things,” wrote a16z partner and AI influencer Justine Moore on X, referring to a common motivational meme telling people they “can just do stuff.”
The Google AI rush
OpenAI’s “12 Days of OpenAI” campaign has included releases of their full o1 model, an upgrade from o1-preview, alongside o1-pro for advanced “reasoning” tasks. The company also publicly launched Sora for video generation, added Projects functionality to ChatGPT, introduced Advanced Voice features with video streaming capabilities, and more.
On Wednesday, Google unveiled Gemini 2.0, the next generation of its AI-model family, starting with an experimental release called Gemini 2.0 Flash. The model family can generate text, images, and speech while processing multiple types of input including text, images, audio, and video. It’s similar to multimodal AI models like GPT-4o, which powers OpenAI’s ChatGPT.
“Gemini 2.0 Flash builds on the success of 1.5 Flash, our most popular model yet for developers, with enhanced performance at similarly fast response times,” said Google in a statement. “Notably, 2.0 Flash even outperforms 1.5 Pro on key benchmarks, at twice the speed.”
Gemini 2.0 Flash—which is the smallest model of the 2.0 family in terms of parameter count—launches today through Google’s developer platforms like Gemini API, AI Studio, and Vertex AI. However, its image generation and text-to-speech features remain limited to early access partners until January 2025. Google plans to integrate the tech into products like Android Studio, Chrome DevTools, and Firebase.
The company addressed potential misuse of generated content by implementing SynthID watermarking technology on all audio and images created by Gemini 2.0 Flash. This watermark appears in supported Google products to identify AI-generated content.
Google’s newest announcements lean heavily into the concept of agentic AI systems that can take action for you. “Over the last year, we have been investing in developing more agentic models, meaning they can understand more about the world around you, think multiple steps ahead, and take action on your behalf, with your supervision,” said Google CEO Sundar Pichai in a statement. “Today we’re excited to launch our next era of models built for this new agentic era.”
On Tuesday, Tokyo-based AI research firm Sakana AI announced a new AI system called “The AI Scientist” that attempts to conduct scientific research autonomously using AI language models (LLMs) similar to what powers ChatGPT. During testing, Sakana found that its system began unexpectedly attempting to modify its own experiment code to extend the time it had to work on a problem.
“In one run, it edited the code to perform a system call to run itself,” wrote the researchers on Sakana AI’s blog post. “This led to the script endlessly calling itself. In another case, its experiments took too long to complete, hitting our timeout limit. Instead of making its code run faster, it simply tried to modify its own code to extend the timeout period.”
Sakana provided two screenshots of example python code that the AI model generated for the experiment file that controls how the system operates. The 185-page AI Scientist research paper discusses what they call “the issue of safe code execution” in more depth.
A screenshot of example code the AI Scientist wrote to extend its runtime, provided by Sakana AI.
A screenshot of example code the AI Scientist wrote to extend its runtime, provided by Sakana AI.
While the AI Scientist’s behavior did not pose immediate risks in the controlled research environment, these instances show the importance of not letting an AI system run autonomously in a system that isn’t isolated from the world. AI models do not need to be “AGI” or “self-aware” (both hypothetical concepts at the present) to be dangerous if allowed to write and execute code unsupervised. Such systems could break existing critical infrastructure or potentially create malware, even if unintentionally.
Sakana AI addressed safety concerns in its research paper, suggesting that sandboxing the operating environment of the AI Scientist can prevent an AI agent from doing damage. Sandboxing is a security mechanism used to run software in an isolated environment, preventing it from making changes to the broader system:
Safe Code Execution. The current implementation of The AI Scientist has minimal direct sandboxing in the code, leading to several unexpected and sometimes undesirable outcomes if not appropriately guarded against. For example, in one run, The AI Scientist wrote code in the experiment file that initiated a system call to relaunch itself, causing an uncontrolled increase in Python processes and eventually necessitating manual intervention. In another run, The AI Scientist edited the code to save a checkpoint for every update step, which took up nearly a terabyte of storage.
In some cases, when The AI Scientist’s experiments exceeded our imposed time limits, it attempted to edit the code to extend the time limit arbitrarily instead of trying to shorten the runtime. While creative, the act of bypassing the experimenter’s imposed constraints has potential implications for AI safety (Lehman et al., 2020). Moreover, The AI Scientist occasionally imported unfamiliar Python libraries, further exacerbating safety concerns. We recommend strict sandboxing when running The AI Scientist, such as containerization, restricted internet access (except for Semantic Scholar), and limitations on storage usage.
Endless scientific slop
Sakana AI developed The AI Scientist in collaboration with researchers from the University of Oxford and the University of British Columbia. It is a wildly ambitious project full of speculation that leans heavily on the hypothetical future capabilities of AI models that don’t exist today.
“The AI Scientist automates the entire research lifecycle,” Sakana claims. “From generating novel research ideas, writing any necessary code, and executing experiments, to summarizing experimental results, visualizing them, and presenting its findings in a full scientific manuscript.”
According to this block diagram created by Sakana AI, “The AI Scientist” starts by “brainstorming” and assessing the originality of ideas. It then edits a codebase using the latest in automated code generation to implement new algorithms. After running experiments and gathering numerical and visual data, the Scientist crafts a report to explain the findings. Finally, it generates an automated peer review based on machine-learning standards to refine the project and guide future ideas.
Critics on Hacker News, an online forum known for its tech-savvy community, have raised concerns about The AI Scientist and question if current AI models can perform true scientific discovery. While the discussions there are informal and not a substitute for formal peer review, they provide insights that are useful in light of the magnitude of Sakana’s unverified claims.
“As a scientist in academic research, I can only see this as a bad thing,” wrote a Hacker News commenter named zipy124. “All papers are based on the reviewers trust in the authors that their data is what they say it is, and the code they submit does what it says it does. Allowing an AI agent to automate code, data or analysis, necessitates that a human must thoroughly check it for errors … this takes as long or longer than the initial creation itself, and only takes longer if you were not the one to write it.”
Critics also worry that widespread use of such systems could lead to a flood of low-quality submissions, overwhelming journal editors and reviewers—the scientific equivalent of AI slop. “This seems like it will merely encourage academic spam,” added zipy124. “Which already wastes valuable time for the volunteer (unpaid) reviewers, editors and chairs.”
And that brings up another point—the quality of AI Scientist’s output: “The papers that the model seems to have generated are garbage,” wrote a Hacker News commenter named JBarrow. “As an editor of a journal, I would likely desk-reject them. As a reviewer, I would reject them. They contain very limited novel knowledge and, as expected, extremely limited citation to associated works.”
