The First AI-Written Paper To Pass Peer Review
April 2, 2025
Cheating. I am not going to bring this topic up.
Humans have taken one small stop towards obsolesce when it comes to writing papers. Sakana AI reports that "The AI Scientist Generates Its First Peer-Reviewed Scientific Publication." This is the first known fully AI-generated paper that passed the same review process that human scientists submit their papers too. Here’s how the paper was written:
"The paper was generated by an improved version of the original AI Scientist, called The AI Scientist-v2. We will be sharing the full details of The AI Scientist-v2 in an upcoming release. This paper was submitted to an ICLR 2025 workshop that agreed to work with our team to conduct an experiment to double-blind review AI-generated manuscripts. We selected this workshop because of its broader scope, challenging researchers (and our AI Scientist) to tackle diverse research topics that address practical limitations of deep learning. The workshop is hosted at ICLR, one of three premier conferences in machine learning and artificial intelligence research, along with NeurIPS and ICML.3
The ICLR leadership and organizers were involved with the project. The paper was blindly submitted to the ICLR review team, although they were told that they might be reviewing AI generated papers.
The AI algorithm was told to research and write about a broad topic. When the process was done, three papers were selected for submission so the review board wouldn’t be overburdened. Here are the results:
“We looked at the generated papers and submitted those we thought were the top 3 (factoring in diversity and quality—We conducted our own detailed analysis of the 3 papers, please read on in our analysis section). Of the 3 papers submitted, two papers did not meet the bar for acceptance. One paper received an average score of 6.33, ranking approximately 45% of all submissions. These scores are higher than many other accepted human-written papers at the workshop, placing the paper above the average acceptance threshold. Specifically, the scores were:
• Rating: 6: Marginally above acceptance threshold
• Rating: 7: Good paper, accept
• Rating: 6: Marginally above acceptance threshold”
The AI Scientist conducted the experiment out of pure scientific curiosity to measure how current AI algorithms compare to human intellect. No problem.
Whitney Grace, April 2, 2025
No Joke: Real Secrecy and Paranoia Are Needed Again
April 1, 2025
No AI. Just a dinobaby sharing an observation about younger managers and their innocence.
In the US and the UK, secrecy and paranoia are chic again. The BBC reported “GCHQ Worker Admits Taking top Secret Data Home.” Ah, a Booz Allen / Snowden type story? The BBC reports:
The court heard that Arshad took his work mobile into a top secret GCHQ area and connected it to work station. He then transferred sensitive data from a secure, top secret computer to the phone before taking it home, it was claimed. Arshad then transferred the data from the phone to a hard drive connected to his personal home computer.
Mr. Snowden used a USB drive. The question is, “What are the bosses doing? Who is watching the logs? Who is checking the video feeds? Who is hiring individuals with some inner need to steal classified information?
But outside phones in a top secret meeting? That sounds like a great idea. I attended a meeting held by a local government agency, and phones and weapons were put in little steel boxes. This outfit was no GHCQ, but the security fellow (a former Marine) knew what he was doing for that local government agency.
A related story addresses paranoia, a mental characteristic which is getting more and more popular among some big dogs.
CNBC reported an interesting approach to staff trust. “Anthropic Announces Updates on Security Safeguards for Its AI Models” reports:
In an earlier version of its responsible scaling policy, Anthropic said it would begin sweeping physical offices for hidden devices as part of a ramped-up security effort.
The most recent update to the firm’s security safeguards adds:
updates to the “responsible scaling” policy for its AI, including defining which of its model safety levels are powerful enough to need additional security safeguards.
The actual explanation is a master piece of clarity. Here’s snippet of what Anthropic actually said in its “Anthropic’s Responsible Scaling Policy” announcement:
The current iteration of our RSP (version 2.1) reflects minor updates clarifying which Capability Thresholds would require enhanced safeguards beyond our current ASL-3 standards.
The Anthropic methods, it seems to me, to include “sweeps” and “compartmentalization.”
Thus, we have two examples of outstanding management:
First, the BBC report implies that personal computing devices can plug in and receive classified information.
And:
Second, CNBC explains that sweeps are not enough. Compartmentalization of systems and methods puts in “cells” who can do what and how.
Andy Grove’s observation popped into my mind. He allegedly rattled off this statement:
Success breeds complacency. Complacency breeds failure. Only the paranoid survive.
Net net: Cyber security is easier to “trust” and “assume”. Real fixes edge into fear and paranoia.
Stephen E Arnold, April 9, 2025
Free AI Sites (Well, Mostly Free Sort of)
April 1, 2025
Dinobaby says, “No smart software involved. That’s for “real” journalists and pundits.
One of my team generated images of French bulldogs. After months of effort, he presented me with a picture of our French bulldog complete with one floppy ear. The image was not free. I pay for the service because free image generation systems work and then degrade because of the costs associated with doing smart software without oodles of cash.
Another person proudly emailed everyone a link to Best AI Websites and the page “Free AI Tools.” The interfaces, functionality, and the outputs vary. The linked Web page is a directory presented with some of that mobile interface zip.l
There are more than 30 tools anyone can try. Here’s what the “directory” interface looks like:
The first click displays the BestFreeAIWebsites’ write up for each “service” or “tool.” Then a direct link to the free AI site is displayed. There is a “submit” button to allow those with a free AI tool to add theirs to the listing. The “add” function is a common feature of Telegram bot and Channel listings.
Here is a selection of the “free” services that are available as of March 28, 2025, in alphabetical order:
- HUUK.ai, a trip planner
- Metavoice at https://studio.themetavoice.xyz/, a “one click voice changer”
- Presentpicker.ai, a service to help a user choose a gift.
- Remaker.ai, a face swap tool
- Yomii.app, a real estate investing assistant
ChatGPT features numerous times in the list of “free” AI tools. Google shows up a couple of times with Bard and Gemini. The majority of the services “wrap” functionality around the big dogs in the LLM space.
Are these services “free”? Our view is that the “free” is a way to get people to give the services a try. If the experience is positive, upgrades are available.
As one of my team worked through the listings, he said, “Most of these services have been available as Telegram bots from other developers.” If he is correct, perhaps Telegram’s AI functions should be included in the listing?
Stephen E Arnold, April 1, 2025
Apple CEO Chases Chinese AI and Phone Sales
March 31, 2025
While the hullabaloo about making stakes in China’s burgeoning market has died down, Big Tech companies still want pieces of the Chinese pie or dumpling would be a better metaphor here. An example of Big Tech wanting to entrench itself in the ChinBaiese market is Apple. Mac Rumors reports that Apple CEO Tim Cook was recently in China and he complimented start-up Deepseek for its AI models. The story, “Apple CEO Tim Cook Praises China’s Deepseek”
While Cook didn’t say he would pursue a partnership with Deepseek, he was impressed with their AI models. He called them excellent, because Deepseek delivers AI models with high performance capabilities that have lower costs and compute requirements. Deepseek’s research has been compared to OpenAI for achieving similar results by using less resources.
When Cook visited China he reportedly made an agreement with Alibaba Group to integrate its Qwen models into Apple Intelligence. There are also rumors that Apple’s speaking with Baidu about providing LLMs for the Chinese market.
Does this mean that Tim Apple hopes he can use Chinese smart tech in the iPhone and make that more appealing to Chinese users? Hmmmm.
Cook conducted more business during his visit:
In addition to his comments on AI, Cook announced plans to expand Apple’s cooperation with the China Development Research Foundation, alongside continued investments in clean energy development. Throughout his visit, Cook posted updates on the Chinese social media platform Weibo, showcasing a range of Apple products being used in classrooms, creative environments, and more.
Cook’s comments mark a continuation of Apple’s intensified focus on the Chinese market at a time when the company is facing declining iPhone shipments and heightened competition from domestic brands. Apple’s smartphone shipments in China are believed to have fallen by 25% year-over-year in the fourth quarter of 2024, while annual shipments dropped 17% to 42.9 million units, placing Apple behind local competitors Vivo and Huawei.”
It’s evident that Apple continues to want a piece of the Chinese dumpling, but also seeks to incorporate Chinese technology into its products. Subtle, Tim Apple, subtle.
Whitney Grace, March 31, 2025
Cypersecurity Pros, Bit of an Issue. You Think?
March 28, 2025
Be aware. A dinobaby wrote this essay. No smart software involved.
I read a research report in the Register titled “MINJA Sneak Attack Poisons AI Models for Other Chatbot Users.” The write up is interesting and, I think, important. The weakness is that the essay does not make explicit that this type of vulnerability can be automated and the outputs used to create the type of weaponized content produced by some intelligence agencies (and PR firms).
The write up provides diagrams and useful detail. For this short blog post, my take on the technique is a manipulation of an LLM’s penchant for adapting to the prompts during a human-interface interaction. If the bad actor crafts misleading information, the outputs can be skewed.
How serious is the behavior in LLMs? In my view, the PR and hype about AI renders the intentional fiddling to a trivial concern. That’s not where the technique nor the implications of its effectiveness belong. Triggering wonky behavior is as easy as mismatching patient data as the article illustrates.
Before one gets too excited about autonomous systems using LLMs to just do it, more attention to the intentional weaponization of LLMs is needed.
Will the AI wizards fix this problem? Sure, someday, but it is an issue that requires time, money, and innovation. We live in an era of marketing. I know I cannot trust most people. Now I know that I can’t trust a MINJA that sneaks into my search or research and delivers a knock out blow.
The Register could have been a bit more energetic in its presentation of this issue. The cited essay does a good job of encouraging bad actors and propagandists to be more diligent in their use of LLMs.
Stephen E Arnold, March 28, 2025
OpenAI and Alleged Environmental Costs: No Problem
March 28, 2025
We know ChatGPT uses an obscene amount of energy and water. But it can be difficult to envision exactly how much. Digg offers some helpful infographics in, "Do You Know How Much Energy ChatGPT Actually Uses?" Writer Darcy Jimenez tells us:
"Since it was first released in 2022, ChatGPT has gained a reputation for being particularly bad for the environment — for example, the GPT-4 model uses as many as 0.14 kilowatt-hours (kWh) generating something as simple as a 100-word email. It can be tricky to fully appreciate the environmental impact of using ChatGPT, though, so the researchers at Business Energy UK made some visualizations to help. Using findings from a 2023 research paper, they calculated the AI chatbot’s estimated water and electricity usage per day, week, month and year, assuming its 200 million weekly users feed it five prompts per day."
See the post for those enlightening graphics. Here are just a few of the astounding statistics:
"Electricity: each day, ChatGPT uses 19.99 million kWh. That’s enough power to charge 4 million phones, or run the Empire State Building for 270 days. … ChatGPT uses a whopping 7.23 billion kWh per year, which is more electricity than the world’s 112 lowest-consumption countries consume over the same period. It’s also enough to power every home in Wyoming for two and a half years."
And:
"Water: The 19.58 million gallons ChatGPT drinks every day could fill a bath for each of Colorado Springs’s 488,664 residents. That amount is also equivalent to everyone in Belgium flushing their toilet at the same time. … In the space of a year, the chatbot uses 7.14 billion gallons of water. That’s enough to fill up the Central Park Reservoir seven times, or power Las Vegas’s Fountains of Bellagio shows for almost 600 years."
Wow. See the write-up for more mind-boggling comparisons. Dolphin lovers and snail darter fans may want to check out the write up.
Cynthia Murrell, March 28, 2025
Programmers: The Way of the Dodo Bird?
March 27, 2025
Another dinobaby blog post. Eight decades and still thrilled when I point out foibles.
Let’s just assume that the US economy is A-OK. One discipline is indispensable now and in the future. What is it? The programmer.
Perhaps not if the information in “Employment for Computer Programmers in the U.S. Has Plummeted to Its Lowest Level Since 1980—Years Before the Internet Existed” is accurate.
The write up states:
There are now fewer computer programmers in the U.S. than there were when Pac-Man was first invented—years before the internet existed as we know it. Computer-programmer employment dropped to its lowest level since 1980, the Washington Post reported, using data from the Current Population Survey from the Bureau of Labor Statistics. There were more than 300,000 computer-programming jobs in 1980. The number peaked above 700,000 during the dot-com boom of the early 2000s but employment opportunities have withered to about half that today. U.S. employment grew nearly 75% in that 45-year period, according to the Post.
What’s interesting is that article makes a classification decision I wasn’t expecting; specifically:
Computer programmers are different from software developers, who liaise between programmers and engineers and design bespoke solutions—a much more diverse set of responsibilities compared to programmers, who mostly carry out the coding work directly. Software development jobs are expected to grow 17% from 2023 to 2033, according to the Bureau of Labor Statistics. The bureau meanwhile projects about a 10% decline in computer programming employment opportunities from 2023 to 2033.
Let’s go with the distinction.
Why are programmers’ jobs disappearing? The write up has the answer:
There has been a 27.5% plummet in the 12-month average of computer-programming employment since about 2023—coinciding with OpenAI’s introduction of ChatGPT the year before. ChatGPT can handle coding tasks without a user needing more detailed knowledge of the code being written. The correlation between the decline of programmer jobs and the rise of AI tools signals to some experts that the burgeoning technology could begin to cost some coding experts their jobs.
Now experts are getting fired? Does that resonate with everyone? Experts.
There is an upside if one indulges in a willing suspension of disbelief. The write up says:
Programmers will be required to perform complicated tasks, Krishna argued, and AI can instead serve to eliminate the simpler, time-consuming tasks those programmers would once need to perform, which would increase productivity and subsequently company performance.
My question, “Did AI contribute to this article?” In my opinion, something is off. It might be dependent on the references to the Bureau of Labor Statistics and “real” newspapers as sources for the numbers. Would a high school debate teacher give the green light to the logic in categorizing and linking those heading for the termination guillotine and those who are on the path to carpet land. The use of AI hype as fact is interesting as well.
I am thrilled to be a dinobaby.
Stephen E Arnold, March 27, 2025
The Chinese AI PR Keeps Flowing
March 27, 2025
Is China moving ahead in the AI race? Some seem to think so. Interesting Engineering reports, "‘World’s First’ Fully Autonomous AI Agent Unveiled in China, Handles Real-World Tasks." Writer Christopher McFadden tells us:
"A group of Chinese software engineers have developed what they have called the ‘world’s first’ fully autonomous artificial intelligence (AI) agent. Called ‘Manus,’ the AI agent can independently perform complex tasks without human guidance. Unlike AI chatbots like ChatGPT, Google’s Gemini, or Grok, which need human input to perform things, Manus can proactively make decisions and complete tasks independently. To this end, the AI agent doesn’t necessarily need to wait for instructions to do something. For example, if a human asks, ‘ Find me an apartment,’ Manus can conduct research, evaluate multiple factors (crime rates, weather, market trends), and provide tailored recommendations."
Apparently, Manus works like a contractor directing their subcontractors. We learn:
"Rather than using just one AI model, Manus operates like an executive managing multiple specialized sub-agents. This allows it to tackle complex, multi-step workflows seamlessly. Moreover, the AI agent can work asynchronously, meaning it completes tasks in the background and notifies users only when results are ready, without constant human supervision. This is a significant development; most AIs have relied heavily on humans to initiate tasks. Manus represents a shift toward fully independent AI, raising exciting possibilities and serious concerns about job displacement and responsibility."
A fully independent AI? Perhaps. If so, the escalated threat to human jobs may be real. Manus has some questioning whether the US is truly the unrivaled leader in the AI space. We shall see if the expectations pan out or are, once again, overblown.
Cynthia Murrell, March 27, 2025
The Future of Programming in an AI Spruik World
March 26, 2025
Software engineers are, reasonably, concerned about losing their jobs to AI. Australian blogger Clinton Boys asks, "How Will LLMs Take Our Jobs?" After reading several posts by programmers using LLMs for side projects, he believes such accounts suggest where we are headed. He writes:
"The consensus seems to be that rather than a side project being some sort of idea you have, then spend a couple of hours on, maybe learn a few things, but quickly get distracted by life or a new side project, you can now just chuck your idea into the model and after a couple of hours of iterating you have a working project. To me, this all seems to point to the fact that we are currently in the middle of a significant paradigm shift, akin to the transition from writing assembly to compiled programming languages. A potential future is unfolding before our eyes in which programmers don’t write in programming languages anymore, but write in natural language, and generative AI handles the grunt work of actually writing the code, the same way a compiler translates your C code into machine instructions."
Perhaps. But then, he ponders, will the job even fit the title of "engineer"? Will the challenges and creative potential many love about this career vanish? And what would they do then? Boys suggests several routes one might take, with the caveat that a realistic path forward would probably blend several of these. He recognizes one could simply give up and choose a different career entirely. An understandable choice, if one can afford to start over. If not, one might join the AI cavalcade by learning how to create LLMs and/or derive value from them. It may also be wise to climb the corporate ladder—managers should be safer longer, Boys expects. Then again one might play ostrich:
"You could also cross your fingers and hope it pans out differently — particularly if, like me you find the vision of the future spruiked by the most bullish LLM proponents a little ghoulish and offensive to our collective humanity."
Always an option, we suppose. I had to look up the Australian term "spruik." According to Wordsmith.org, it means "to make an elaborate speech, especially to attract customers." Fitting. Finally, Boys says, one could bet on software connoisseurs of the future. Much as some now pay more for hand-made pastries or small-batch IPAs, some clients may be willing to shell out for software crafted the old-fashioned way. One can hope.
Cynthia Murrell, March 26, 2025
Old School Search: Scrunch Can Help You
March 25, 2025
Google, DuckDuckGo, Bing, and other search engines have incorporated AI into their search algorithms. AI, however, remains regulated to generative text and chatbots. It’s also doing very little to assist companies with their Web presences Tech Crunch shares how one startup wants to change that: “Scrunch AI Is Helping Companies Stand Out In AI Search.”
Scrunch AI developed a platform that assists companies with auditing and optimizing how their appear on AI search platforms. The platform shows how a company’s online information interacts with AI Web crawlers. Scrunch also funds inaccuracies and gaps in information.
The CEO and co-founder of Crunch AI Chris Andrew said he got the idea for the platform when he realized that he expected ChatGPT to do the browsing for him. He shared the idea with CMOs who noticed that their companies received high-quality traffic from AI search engines. The rub, however, was that the companies received different results from different platforms.
While there are companies that concentrate o this task, he says Scrunch goes further than then:
“Andrew thinks his startup stands out thanks to its focus on the customer journey as opposed to just how a brand shows up in initial search results. He feels the company is also taking it a step further by not just focusing on search results by a human through an AI search engine, but rather on searches performed by AI agents. ‘I think people were like, ‘How do we use AI to make our website better?’ And my mindset was like, ‘Your website’s going to need to be for an agent or crawler in the future,’” Andrew said. ‘That theory has kind of really played out with our customer base at the enterprise level saying our brand is no longer what we say it is. It’s what ChatGPT, Gemini, Siri, Google AI Overviews say it is.’”
Consistency and accuracy is important in this digital age. Andrew has a great idea but will Scrunch optimize search engine AI or will it generate AI slop?
Whitney Grace, March 25, 2025
-
- Subscribe to Beyond Search
Feature archive
News archive
-
