Is There a Problem with AI Detection Software?
July 1, 2024
Of course not.
But colleges and universities are struggling to contain AI-enabled cheating. Sadly, it seems the easiest solution is tragically flawed. Times Higher Education considers, “Is it Time to Turn Off AI Detectors?” The post shares a portion of the new book, “Teaching with AI: A Practical Guide to a New Era of Human Learning” by José Antonio Bowen and C. Edward Watson. The excerpt begins by looking at the problem:
“The University of Pennsylvania’s annual disciplinary report found a seven-fold (!) increase in cases of ‘unfair advantage over fellow students’, which included ‘using ChatGPT or Chegg’. But Quizlet reported that 73 per cent of students (of 1,000 students, aged 14 to 22 in June 2023) said that AI helped them ‘better understand material’. Watch almost any Grammarly ad (ubiquitous on TikTok) and ask first, if you think clicking on ‘get citation‘ or ‘paraphrase‘ is cheating. Second, do you think students might be confused?”
Probably. Some universities are not exactly clear on what is cheating and what is permitted usage of AI tools. At the same time, a recent study found 51 percent of students will keep using them even if they are banned. The boost to their GPAs is just too tempting. Schools’ urge to fight fire with fire is understandable, but detection tools are far from perfect. We learn:
“AI detectors are already having to revise claims. Turnitin initially claimed a 1 per cent false-positive rate but revised that to 4 per cent later in 2023. That was enough for many institutions, including Vanderbilt, Michigan State and others, to turn off Turnitin’s AI detection software, but not everyone followed their lead. Detectors vary considerably in their accuracy and rate of false positives. One study looked at 14 different detectors and found that five of the 14 were only 50 per cent accurate or worse, but four of them (CheckforAI, Winston AI, GPT-2 Output and Turnitin) missed only one of the 18 AI-written samples. Detectors are not all equal, but the best are better than faculty at identifying AI writing.”
But is that ability is worth the false positives? One percent may seem small, but to those students it can mean an end to their careers before they even begin. For institutions that do not want to risk false accusations, the authors suggest several alternatives that seem to make a difference. They advise instructors to discuss the importance of academic integrity at the beginning of the course and again as the semester progresses. Demonstrating how well detection tools work can also have an impact. Literally quizzing students on the school’s AI policies, definitions, and consequences can minimize accidental offenses. Schools could also afford students some wiggle room: allow them to withdraw submissions and take the zero if they have second thoughts. Finally, the authors suggest schools normalize asking for help. If students get stuck, they should feel they can turn to a human instead of AI.
Cynthia Murrell, July 1, 2024
Some Tension in the Datasphere about Artificial Intelligence
June 28, 2024
This essay is the work of a dinobaby. Unlike some folks, no smart software improved my native ineptness.
I generally try to avoid profanity in this blog. I am mindful of Google’s stopwords. I know there are filters running to protect those younger than I from frisky and inappropriate language. Therefore, I will cite the two articles and then convert the profanity to a suitably sanitized form.
The first write up is “I Will F…ing Piledrive You If You Mention AI Again”. Sorry, like many other high-technology professionals I prevaricated and dissembled. I have edited the F word to be less superficially offensive. (One simply cannot trust high-technology types, can you? I am not Thomson Reuters obviously.) The premise of this write up is that smart software is over-hyped. Here’s a passage I found interesting:
Unless you are one of a tiny handful of businesses who know exactly what they’re going to use AI for, you do not need AI for anything – or rather, you do not need to do anything to reap the benefits. Artificial intelligence, as it exists and is useful now, is probably already baked into your businesses software supply chain. Your managed security provider is probably using some algorithms baked up in a lab software to detect anomalous traffic, and here’s a secret, they didn’t do much AI work either, they bought software from the tiny sector of the market that actually does need to do employ data scientists.
I will leave it to you to ponder the wisdom of these words. I, for instance, do not know exactly what I am going to do until I do something, fiddle with it, and either change it up or trash it. You and most AI enthusiasts are probably different. That’s good. I envy your certitude. The author of the first essay is not gentle; he wants to piledrive you if you talk about smart software. I do not advocate violence under any circumstances. I can tolerate baloney about smart software. The piledriver person has hate in his heart. You have been warned.
The second write up is “ChatGPT Is Bullsh*t,” and it is an article published in SpringerLink, not a personal blog. Yep, bullsh*t as a term in an academic paper. Keep in mind, please, that Stanford University’s president and some Harvard wizards engaged in the bullsh*t business as part of their alleged making up data. Who needs AI when humans are perfectly capable of hallucinating, but I digress?
I noted this passage in the academic write up:
So perhaps we should, strictly, say not that ChatGPT is bullshit but that it outputs bullshit in a way that goes beyond being simply a vector of bullshit: it does not and cannot care about the truth of its output, and the person using it does so not to convey truth or falsehood but rather to convince the hearer that the text was written by a interested and attentive agent.
Please, read the 10 page research article about bullsh*t, soft bullsh*t, and hard bullsh*t. Form your own opinion.
I have now set the stage for some observations (probably unwanted and deeply disturbing to some in the smart software game).
- Artificial intelligence is a new big thing, and the hyperbole, misdirection, and outright lying like my saying I would use forbidden language in this essay irrelevant. The object of the new big thing is to make money, get power, maybe become an influencer on TikTok.
- The technology which seems to have flowered in January 2023 when Microsoft said, “We love OpenAI. It’s a better Clippy.” The problem is that it is now June 2024 and the advances have been slow and steady. This means that after a half century of research, the AI revolution is working hard to keep the hypemobile in gear. PR is quick; smart software improvement less speedy.
- The ripples the new big thing has sent across the datasphere attenuate the farther one is from the January 2023 marketing announcement. AI fatigue is now a thing. I think the hostility is likely to increase because real people are going to lose their jobs. Idle hands are the devil’s playthings. Excitement looms.
Net net: I think the profanity reveals the deep disgust some pundits and experts have for smart software, the companies pushing silver bullets into an old and rusty firearm, and an instinctual fear of the economic disruption the new big thing will cause. Exciting stuff. Oh, I am not stating a falsehood.
Stephen E Arnold, June 23, 2024
Can the Bezos Bulldozer Crush Temu, Shein, Regulators, and AI?
June 27, 2024
This essay is the work of a dumb dinobaby. No smart software required.
The question, to be fair, should be, “Can the Bezos-less bulldozer crush Temu, Shein, Regulators, Subscriptions to Alexa, and AI?” The article, which appeared in the “real” news online service Venture Beat, presents an argument suggesting that the answer is, “Yes! Absolutely.”
Thanks MSFT Copilot. Good bulldozer.
The write up “AWS AI Takeover: 5 Cloud-Winning Plays They’re [sic] Using to Dominate the Market” depends upon an Amazon Big Dog named Matt Wood, VP of AI products at AWS. The article strikes me as something drafted by a small group at Amazon and then polished to PR perfection. The reasons the bulldozer will crush Google, Microsoft, Hewlett Packard’s on-premises play, and the keep-on-searching IBM Watson, among others, are:
- Covering the numbers or logo of the AI companies in the “game”; for example, Anthropic, AI21 Labs, and other whale players
- Hitting up its partners, customers, and friends to get support for the Amazon AI wonderfulness
- Engineering AI to be itty bitty pieces one can use to build a giant AI solution capable of dominating D&B industry sectors like banking, energy, commodities, and any other multi-billion sector one cares to name
- Skipping the Google folly of dealing with consumers. Amazon wants the really big contracts with really big companies, government agencies, and non-governmental organizations.
- Amazon is just better at security. Those leaky S3 buckets are not Amazon’s problem. The customers failed to use Amazon’s stellar security tools.
Did these five points convince you?
If you did not embrace the spirit of the bulldozer, the Venture Beat article states:
Make no mistake, fellow nerds. AWS is playing a long game here. They’re not interested in winning the next AI benchmark or topping the leaderboard in the latest Kaggle competition. They’re building the platform that will power the AI applications of tomorrow, and they plan to power all of them. AWS isn’t just building the infrastructure, they’re becoming the operating system for AI itself.
Convinced yet? Well, okay. I am not on the bulldozer yet. I do hear its engine roaring and I smell the no-longer-green emissions from the bulldozer’s data centers. Also, I am not sure the Google, IBM, and Microsoft are ready to roll over and let the bulldozer crush them into the former rain forest’s red soil. I recall researching Sagemaker which had some AI-type jargon applied to that “smart” service. Ah, you don’t know Sagemaker? Yeah. Too bad.
The rather positive leaning Amazon write up points out that as nifty as those five points about Amazon’s supremacy in the AI jungle, the company has vision. Okay, it is not the customer first idea from 1998 or so. But it is interesting. Amazon will have infrastructure. Amazon will provide model access. (I want to ask, “For how long?” but I won’t.), and Amazon will have app development.
The article includes a table providing detail about these three legs of the stool in the bulldozer’s cabin. There is also a run down of Amazon’s recent media and prospect directed announcements. Too bad the article does not include hyperlinks to these documents. Oh, well.
And after about 3,300 words about Amazon, the article includes about 260 words about Microsoft and Google. That’s a good balance. Too bad IBM. You did not make the cut. And HP? Nope. You did not get an “Also participated” certificate.
Net net: Quite a document. And no mention of Sagemaker. The Bezos-less bulldozer just smashes forward. Success is in crushing. Keep at it. And that “they” in the Venture Beat article title: Shouldn’t “they” be an “it”?
Stephen E Arnold, June 27, 2024
Nerd Flame War: AI AI AI
June 27, 2024
The Internet is built on trolls and their boorish behavior. The worst of the trolls are self-confessed “experts” on anything. Every online community has their loitering trolls and tech enthusiasts aren’t any different. In the old days of Internet lore, online verbal battles were dubbed “flame wars” and XDA-Developers reports that OpenAI started one: “AI Has Thrown Stack Overflow Into Civil War.”
A huge argument in AI development is online content being harvested for large language models (LLMs) to train algorithms. Writers and artists were rightly upset were used to train image and writing algorithms. OpenAI recently partnered with Stack Overflow to collect data and the users aren’t happy. Stack Overflow is a renowned tech support community for sysadmin, developers, and programmers. Stack Overflow even brags that it is world’s largest developer community.
Stack Overflow users are angry, because they weren’t ask permission to use their content for AI training models and they don’t like the platform’s response to their protests. Users are deleting their posts or altering them to display correct information. In response, Stack Overflow is restoring deleted and incorrect information, temporarily suspending users who delete content, and hiding behind the terms of service. The entire situation is explained here:
“Delving into discussion online about OpenAI and Stack Overflow’s partnership, there’s plenty to unpack. The level of hostility towards Stack Overflow varies, with some users seeing their answers as being posted online without conditions – effectively free for all to use, and Stack Overflow granting OpenAI access to that data as no great betrayal. These users might argue that they’ve posted their answers for the betterment of everyone’s knowledge, and don’t place any conditions on its use, similar to a highly permissive open source license.
Other users are irked that Stack Overflow is providing access to an open-resource to a company using it to build closed-source products, which won’t necessarily better all users (and may even replace the site they were originally posted on.) Despite OpenAI’s stated ambition, there is no guarantee that Stack Overflow will remain freely accessible in perpetuity, or that access to any AIs trained on this data will be free to the users who contributed to it.”
Reddit and other online communities are facing the same problems. LLMs are made from Stack Overflow and Reddit to train generative AI algorithms like ChatGPT. OpenAI’s ChatGPT is regarded as overblown because it continues to fail multiple tests. We know, however, that generative AI will improve with time. We also know that people will use the easiest solution and generative AI chatbots will become those tools. It’s easier to verbally ask or write a question than searching.
Whitney Grace, June 27, 2024
Two EU Firms Unite in Pursuit of AI Sovereignty
June 25, 2024
Europe would like to get out from under the sway of North American tech firms. This is unsurprising, given how differently the EU views issues like citizen privacy. Then there are the economic incentives of localizing infrastructure, data, workforce, and business networks. Now, two generative AI firms are uniting with that goal in mind. The Next Web reveals, “European AI Leaders Aleph Alpha and Silo Ink Deal to Deliver ‘Sovereign AI’.” Writer Thomas Macaulay reports:
“Germany’s Aleph Alpha and Finland’s Silo AI announced the partnership [on June 13, 2024]. The duo plan to create a ‘one-stop-solution’ for European industrial firms exploring generative AI. Their collaboration brings together distinctive expertise. Aleph Alpha has been described a European rival to OpenAI, but with a stronger focus on data protection, security, and transparency. The company also claims to operate Europe’s fastest commercial AI data center. Founded in 2019, the firm has become Germany’s leading AI startup. In November, it raised $500mn in a funding round backed by Bosch, SAP, and Hewlett Packard Enterprise. Silo AI, meanwhile, calls itself ‘Europe’s largest private AI lab.’ The Helsinki-based startup provides custom LLMs through a SaaS subscription. Use cases range from smart devices and cities to autonomous vehicles and industry 4.0. Silo also specializes in building LLMs for low-resource languages, which lack the linguistic data typically needed to train AI models. By the end of this year, the company plans to cover every official EU language.”
Both Aleph Alpha CEO Jonas Andrulis and Silo AI CEO Peter Sarlin enthusiastically advocate European AI sovereignty. Will the partnership strengthen their mutual cause?
Cynthia Murrell, June 25, 2024
A Discernment Challenge for Those Who Are Dull Normal
June 24, 2024
This essay is the work of a dinobaby. Unlike some folks, no smart software improved my native ineptness.
Techradar, an online information service, published “Ahead of GPT-5 Launch, Another Test Shows That People Cannot Distinguish ChatGPT from a Human in a Conversation Test — Is It a Watershed Moment for AI?” The headline implies “change everything” rhetoric, but that is routine AI jargon-hype.
Once again, academics who are unable to land a job in a “real” smart software company studied the work of their former colleagues who make a lot more money than those teaching do. Well, what do academic researchers do when they are not sitting in the student union or the snack area in the lab whilst waiting for a graduate student to finish a task? In my experience, some think about their CVs or résumés. Others ponder the flaws in a commercial or allegedly commercial product or service.
A young shopper explains that the outputs of egg laying chickens share a similarity. Insightful observation from a dumb carp. Thanks, MSFT Copilot. How’s that Recall project coming along?
The write up reports:
The Department of Cognitive Science at UC San Diego decided to see how modern AI systems fared and evaluated ELIZA (a simple rules-based chatbot from the 1960’s included as a baseline in the experiment), GPT-3.5, and GPT-4 in a controlled Turing Test. Participants had a five-minute conversation with either a human or an AI and then had to decide whether their conversation partner was human.
Here’s the research set up:
In the study, 500 participants were assigned to one of five groups. They engaged in a conversation with either a human or one of the three AI systems. The game interface resembled a typical messaging app. After five minutes, participants judged whether they believed their conversation partner was human or AI and provided reasons for their decisions.
And what did the intrepid academics find? Factoids that will get them a job at a Perplexity-type of company? Information that will put smart software into focus for the elected officials writing draft rules and laws to prevent AI from making The Terminator come true?
The results were interesting. GPT-4 was identified as human 54% of the time, ahead of GPT-3.5 (50%), with both significantly outperforming ELIZA (22%) but lagging behind actual humans (67%). Participants were no better than chance at identifying GPT-4 as AI, indicating that current AI systems can deceive people into believing they are human.
What does this mean for those labeled dull normal, a nifty term applied to some lucky people taking IQ tests. I wanted to be a dull normal, but I was able to score in the lowest possible quartile. I think it was called dumb carp. Yes!
Several observations to disrupt your clear thinking about smart software and research into how the hot dogs are made:
- The smart software seems to have stalled. Our tests of You.com which allows one to select which object models parrots information, it is tough to differentiate the outputs. Cut from the same transformer cloth maybe?
- Those judging, differentiating, and testing smart software outputs can discern differences if they are way above dull normal or my classification dumb carp. This means that indexing systems, people, and “new” models will be bamboozled into thinking what’s incorrect is a-okay. So much for the informed citizen.
- Will the next innovation in smart software revolutionize something? Yep, some lucky investors.
Net net: Confusion ahead for those like me: Dumb carp. Dull normals may be flummoxed. But those super-brainy folks have a chance to rule the world. Bust out the party hats and little horns.
Stephen E Arnold, June 24, 2024
Ad Hominem Attack: A Revived Rhetorical Form
June 24, 2024
This essay is the work of a dinobaby. Unlike some folks, no smart software improved my native ineptness.
I remember my high school debate coach telling my partner Nick G. (I have forgotten the budding prosecutor’s name, sorry) you should not attack the character of our opponents. Nick G. had interacted with Bill W. on the basketball court in an end-of-year regional game. Nick G., as I recall got a bloody nose, and Bill W. was thrown out of the basketball game. When fisticuffs ensued, I thanked my lucky stars I was a hopeless athlete. Give me the library, a debate topic, a pile of notecards, and I was good to go. Nick G. included in his rebuttal statement comments about the character of Bill W. When the judge rendered a result and his comments, Nick G. was singled out as being wildly inappropriate. After the humiliating defeat, the coach explained that an ad hominem argument is not appropriate for 15-year-olds. Nick G.’s attitude was, “I told the truth.” As Nick G. learned, the truth is not what wins debate tournaments or life in some cases.
I thought about ad hominem arguments as I read “Silicon Valley’s False Prophet.” This essay reminded me of the essay by the same author titled “The Man Who Killed Google Search.” I must admit the rhetorical trope is repeatable. Furthermore it can be applied to an individual who may be clueless about how selling advertising nuked relevance (or what was left of it) at the Google and to the dealing making of a person whom I call Sam AI-Man. Who knows? Maybe other authors will emulate these two essays, and a new Silicon Valley genre may emerge ready for the real wordsmiths and pooh-bahs of Silicon Valley to crank out a hit piece every couple of days.
To the essay at hand: The false profit is the former partner of Elon Musk and the on-again-off-again-on-again Big Dog at OpenAI. That’s an outfit where “open” means closed, and closed means open to the likes of Apple. The main idea, I think, is that AI sucks and Sam AI-Man continues to beat the drum for a technology that is likely to be headed for a correction. In Silicon Valley speak, the bubble will burst. It is, I surmise, Mr. AI-man’s fault.
The essay explains:
Sam Altman, however, exists in a category of his own. There are many, many, many examples of him saying that OpenAI — or AI more broadly — will do something it can’t and likely won’t, and it being meekly accepted by the Fourth Estate without any real pushback. There are more still of him framing the limits of the present reality as a positive — like when, in a fireside sitdown with
1980s used car salesmanSalesforce CEO Marc Benioff, Altman proclaimed that AI hallucinations (when an LLM asserts something untrue as fact, because AI doesn’t know anything) are a feature, not a bug, and rather than being treated as some kind of fundamental limitation, should be regarded as a form of creative expression.
I understand. Salesperson. Quite a unicorn in Silicon Valley. I mean when I worked there I would encounter hyperbole artists every few minutes. Yeah, Silicon Valley. Anchored in reality, minimum viable products, and lots of hanky pinky.
The essay provides a bit of information about the background of Mr. AI-Man:
When you strip away his ability to convince people that he’s smart, Altman had actually done very little — he was a college dropout with a failing-then-failed startup, one where employees tried to get him fired twice.
If true, that takes some doing. Employees tried to get the false prophet fired twice. In olden times, burning at the stake might have been an option. Now it is just move on to another venture. Progress.
The essay does provide some insight into Sam AI-Man’s core competency:
Altman is adept at using connections to make new connections, in finding ways to make others owe him favors, in saying the right thing at the right time when he knew that nobody would think about it too hard. Altman was early on Stripe, and Reddit, and Airbnb — all seemingly-brilliant moments in the life of a man who had many things handed to him, who knew how to look and sound to get put in the room and to get the capital to make his next move. It’s easy to conflate investment returns with intellectual capital, even though the truth is that people liked Altman enough to give him the opportunity to be rich, and he took it.
I cannot figure out if the author envies Sam AI-Man, reviles him for being clever (a key attribute in some high-technology outfits), or genuinely perceives Mr. AI-Man as the first cousin to Beelzebub. Whatever the motivation, I find the phoenix-like rising of the ad hominem attack a refreshing change from the entitled pooh-bahism of some folks writing about technology.
The only problem: I think it is unlikely that the author will be hired by OpenAI. Chance blown.
Stephen E Arnold, June 24, 2024
Chasing a Folly: Identifying AI Content
June 24, 2024
As are other academic publishers, Springer Nature Group is plagued by fake papers. Now the company announces, “Springer Nature Unveils Two New AI Tools to Protect Research Integrity.” How effective the tools are remains to be proven, but at least the company is making an effort. The press release describes text-checker Geppetto and image-analysis tool SnappShot. We learn:
“Geppetto works by dividing the paper up into sections and uses its own algorithms to check the consistency of the text in each section. The sections are then given a score based on the probability that the text in them has been AI generated. The higher the score, the greater the probability of there being problems, initiating a human check by Springer Nature staff. Geppetto is already responsible for identifying hundreds of fake papers soon after submission, preventing them from being published – and from taking up editors’ and peer reviewers’ valuable time.
SnappShot, also developed in-house, is an AI-assisted image integrity analysis tool. Currently used to analyze PDF files containing gel and blot images and look for duplications in those image types – another known integrity problem within the industry – this will be expanded to cover additional image types and integrity problems and speed up checks on papers.”
Springer Nature’s Chris Graf emphasizes the importance of research integrity and vows to continue developing and improving in-house tools. To that end, we learn, the company is still growing its fraud-detection team. The post points out Springer Nature is a contributing member of the STM Integrity Hub.
Based in Berlin, Springer Nature was formed in 2015 through the combination of Nature Publishing Group, Macmillan Education, and Springer Science+Business Media. A few of its noteworthy publications include Scientific American, Nature, and this collection of Biology, Clinical Medicine, and Health journals.
Cynthia Murrell, June 24, 2024
The Key to Success at McKinsey & Company: The 2024 Truth Is Out!
June 21, 2024
This essay is the work of a dinobaby. Unlike some folks, no smart software improved my native ineptness.
When I was working at a “real” company, I wanted to labor in the vineyards of a big-time, blue-chip consulting firm. I achieved that goal and, after a suitable period of time in the penal colony, I escaped to a client. I made it out, unscathed, and entered a more interesting, less nutso working life. When the “truth” about big-time, blue-chip consulting firms appears in public sources, I scan the information. Most of it is baloney; for example, the yip yap about McKinsey and its advice pertaining to addictive synthetics. Hey, stuff happens when one is objective. “McKinsey Exec Tells Summer Interns That Learning to Ask AI the Right Questions Is the Key to Success” contains some information which I find quite surprising. First, I don’t know if the factoids in the write up are accurate or if they are the off-the-cuff baloney recruiters regularly present to potential 60-hour-a-week knowledge worker serfs or if the person has a streaming video connection to the McKinsey managing partner’s work-from-the-resort office.
Let’s assume the information is correct and consider some of its implications. An intern is a no-pay or low-pay job for students from the right institutions, the right background, or the right connections. The idea is that associates (one step above the no-pay serf) and partners (the set for life if you don’t die of heart failure crowd) can observe, mentor, and judge these field laborers. The write up states:
Standing out in a summer internship these days boils down to one thing — learning to talk to AI. At least, that’s the advice McKinsey’s chief client officer, Liz Hilton Segel, gave one eager intern at the firm. “My advice to her was to be an outstanding prompt engineer,” Hilton Segel told The Wall Street Journal.
But what about grades? What about my family’s connections to industry, elected officials, and a supreme court judge? What about my background scented with old money, sheepskin from prestigious universities, and a Nobel Prize awarded a relative 50 years ago? These questions, its seems, may no longer be relevant. AI is coming to the blue-chip consulting game, and the old-school markers of building big revenues may not longer matter.
AI matters. Despite McKinsey’s 11-month effort, the firm has produced Lilli. The smart systems, despite fits and starts, has delivered results; that is, a payoff, cash money, engagement opportunities. The write up says:
Lilli’s purpose is to aggregate the firm’s knowledge and capabilities so that employees can spend more time engaging with clients, Erik Roth, a senior partner at McKinsey who oversaw Lili’s development, said last year in a press release announcing the tool.
And the proof? I learned:
“We’ve [McKinsey humanoids] answered over 3 million prompts and add about 120,000 prompts per week,” he [Erik Roth] said. “We are saving on average up to 30% of a consultants’ time that they can reallocate to spend more time with their clients instead of spending more time analyzing things.”
Thus, the future of success is to learn to use Lilli. I am surprised that McKinsey does not sell internships, possibly using a Ticketmaster-type system.
Several observations:
- As Lilli gets better or is replaced by a more cost efficient system, interns and newly hired professionals will be replaced by smart software.
- McKinsey and other blue-chip outfits will embrace smart software because it can sell what the firm learns to its clients. AI becomes a Petri dish for finding marketable information.
- The hallucinative functions of smart software just create an opportunity for McKinsey and other blue-chip firms to sell their surviving professionals at a more inflated fee. Why fail and lose money? Just pay the consulting firm, sidestep the stupidity tax, and crush those competitors to whom the consulting firms sell the cookie cutter knowledge.
Net net: Blue-chip firms survived the threat from gig consultants and the Gerson Lehrman-type challenge. Now McKinsey is positioning itself to create a no-expectation environment for new hires, cut costs, and increase billing rates for the consultants at the top of the pyramid. Forget opioids. Go AI.
Stephen E Arnold, June 21, 2024
Can Anthropic Break Into the AI Black Box?
June 20, 2024
The inner workings of large language models have famously been a mystery, even to their creators. That is a problem for those who would like transparency around pivotal AI systems. Now, however, Anthropic may have found the solution. Time reports, “No One Truly Knows Bow AI Systems Work. A New Discovery Could Change That.” If the method pans out, this will be perfect for congressional hearings and anti trust testimony. Reporter Billy Perrigo writes:
“Researchers developed a technique for essentially scanning the ‘brain’ of an AI model, allowing them to identify collections of neurons—called ‘features’—corresponding to different concepts. And for the first time, they successfully used this technique on a frontier large language model, Anthropic’s Claude Sonnet, the lab’s second-most powerful system, .In one example, Anthropic researchers discovered a feature inside Claude representing the concept of ‘unsafe code.’ By stimulating those neurons, they could get Claude to generate code containing a bug that could be exploited to create a security vulnerability. But by suppressing the neurons, the researchers found, Claude would generate harmless code. The findings could have big implications for the safety of both present and future AI systems. The researchers found millions of features inside Claude, including some representing bias, fraudulent activity, toxic speech, and manipulative behavior. And they discovered that by suppressing each of these collections of neurons, they could alter the model’s behavior. As well as helping to address current risks, the technique could also help with more speculative ones.”
The researchers hope their method will replace “red-teaming,” where developers chat with AI systems in order to uncover toxic or dangerous traits. On the as-of-yet theoretical chance an AI gains the capacity to deceive its creators, the more direct method would be preferred.
A happy side effect of the method could be better security. Anthropic states being able to directly manipulate AI features may allow developers to head off AI jailbreaks. The research is still in the early stages, but Anthropic is singing an optimistic tune.
Cynthia Murrell, June 20, 2024
-
- Subscribe to Beyond Search
Feature archive
News archive
-
