Apple AI Is AImless: Better Than Fire, Ready AIm
May 16, 2025
Apple’s Problems Rebuilding Siri
Apple is a dramatist worthy of reality TV. According to MSN, Apple’s leaders are fighting each other says the article, “New Siri Report Reveals Epic Dysfunction Within Apple — But There’s Hope.” There’s so many issues with Apple’s leaders that Siri 2.0 is delayed until 2026.
Managerial styles and backroom ambitions clashed within Apple’s teams. John Giannandrea heads Siri and has since 2018. He was hired to lead Siri and an AI group. Siri engineers claim they are treated like second class citizens. Their situation worsened when Craig Federighi’s software team released features and updates.
The two leaders are very different:
“Federighi was placed in charge of the Siri overhaul in March, alongside his number two Mike Rockwell — who created the Apple Vision Pro headset— as Apple attempts to revive its Siri revamp. The difference between Giannandrea and Federighi appears to be the difference between the tortoise and the hare. John is allegedly more of a listener and slow mover who lets those underneath him take charge of the work, especially his number two Robby Walker. He reportedly preferred incremental updates and was repeatedly cited as a problem with Siri development. Meanwhile, Federighi is described as brash and quick but very efficient and knowledgeable. Supposedly, Giannandrea’s “relaxed culture” lead to other engineers dubbing his AI team: AIMLess.”
The two teams are at each other’s throats. Projects are getting done but they’re arguing over the means of how to do them. Siri 2.0 is caught in the crossfire like a child of divorce. The teams need to put their egos aside or someone in charge of both needs to make them play nicely.
Whitney Grace, May 16, 2025
Retail Fraud Should Be Spelled RetAIl Fraud
May 16, 2025
As brick-and-mortar stores approach extinction and nearly all shopping migrates to the Web, AI introduces new vulnerabilities to the marketplace. Shocking, we know. Cyber Security Intelligence reports, “ChatGPT’s Image Generation Could Be Driving Retail Fraud.” We learn:
“The latest AI image generators can create images that look like real photographs as well as imagery from simple text prompts with incredible accuracy. It can reproduce documents with precisely matching formatting, official logos, accurate timestamps, and even realistic barcodes or QR codes. In the hands of fraudsters, these tools can be used to commit ‘return fraud’ by creating convincing fake receipts and proof-of-purchase documentation.”
But wait, there is more. The post continues:
“Fake proof of purchase documentation can be used to claim warranty service for products that are out of warranty or purchased through unauthorised channels. Fraudsters could also generate fake receipts showing purchases at higher values than was actually paid for – then requesting refunds to gift cards for the inflated amount. Internal threats also exist too, as employees can create fake expense receipts for reimbursement. This is particularly damaging for businesses with less sophisticated verification processes in place. Perhaps the scenario most concerning of all is that these tools can enable scammers to generate convincing payment confirmations or shipping notices as part of larger social engineering attacks.”
Also of concern is the increased inconvenience to customers as sites beef up their verification processes. After all, the write-up notes, The National Retail Federation found 70% of customers say a positive return experience makes them more likely to revisit a seller.
So what is a retail site to do? Well, author Doriel Abrahams is part of Forter, a company that uses AI to protect online sellers from fraud. Naturally, he suggests using a platform like his firm’s to find suspicious patterns without hindering legit customers too much. Is more AI the solution? We are not certain. If one were to go down that route, though, one should probably compare multiple options.
Cynthia Murrell, May 16, 2025
Complexity: Good Enough Is Now the Best Some Can Do at Google-
May 15, 2025
![dino-orange_thumb_thumb_thumb_thumb_[1]_thumb_thumb](https://www.arnoldit.com/wordpress/wp-content/uploads/2025/05/dino-orange_thumb_thumb_thumb_thumb_1_thumb_thumb-1.gif "dino-orange_thumb_thumb_thumb_thumb_[1]_thumb_thumb")
No AI, just the dinobaby expressing his opinions to Zillennials.
I read a post called “Working on Complex Systems: What I Learned Working at Google.” The write up is a thoughtful checklist of insights, lessons, and Gregorian engineering chants a “coder” learned in the online advertising company. I want to point out that I admire the amount of money and power the Google has amassed from its reinvention of the GoTo-Overture-Yahoo advertising approach.
A Silicon Valley executive looks at past due invoices. The government has ordered the company to be broken up and levied large fines for improper behavior in the marketplace. Thanks, ChatGPT. Definitely good enough.
The essay in The Coder Cafe presents an engineer’s learnings after Google began to develop products and services tangential to search hegemony, selling ads, and shaping information flows.
The approach is to differentiate complexity from complicated systems. What is interesting about the checklists is that one hearkens back to the way Google used to work in the Backrub and early pre-advertising days at Google. Let’s focus on complex because that illuminates where Google wants to direct its business, its professionals, its users, and the pesky thicket of regulators who bedevil the Google 24×7.
Here’s the list of characteristics of complex systems. Keep in mind that “systems” means software, programming, algorithms, and the gizmos required to make the non-fungible work, mostly.
- Emergent behavior
- Delayed consequences
- Optimization (local optimization versus global optimization)
- Hysteresis (I think this is cultural momentum or path dependent actions)
- Nonlinearity
Each of these is a study area for people at the Santa Fe Institute. I have on my desk a copy of The Origins of Order: Self-Organization and Selection in Evolution and the shorter Reinventing the Sacred, both by Stuart A. Kauffman. As a point of reference Origins is 700 pages and Reinventing about 300. Each of the cited articles five topics gets attention.
The context of emergent behavior in human- and probably some machine- created code is that it is capable of producing “complex systems.” Dr. Kauffman does a very good job of demonstrating how quite simple methods yield emergent behavior. Instead of a mess or a nice tidy solution, there is considerable activity at the boundaries of complexity and stability. Emergence seems to be associated with these boundary conditions: A little bit of chaos, a little bit of stability.
The other four items in the list are optimization. Dr. Kauffman points out is a consequence of the simple decisions which take place in the micro and macroscopic world. Non-linearity is a feature of emergent systems. The long-term consequences of certain emergent behavior can be difficult to predict. Finally, the notion of momentum keeps some actions or reactions in place through time units.
What the essay reveals, in my opinion, that:
- Google’s work environment is positioned as a fundamental force. Dr. Kauffman and his colleagues at the Santa Fe Institute may find some similarities between the Google and the mathematical world at the research institute. Google wants to be the prime mover; the Santa Fe Institute wants to understand, explain, and make useful its work.
- The lingo of the cited essay suggests that Google is anchored in the boundary between chaos and order. Thus, Google’s activities are in effect trials and errors intended to allow Google to adapt and survive in its environment. In short, Google is a fundamental force.
- The “leadership” of Google does not lead; leadership is given over to the rules or laws of emergence as described by Dr. Kauffman and his colleagues at the Santa Fe Institute.
Net net: Google cannot produce good products. Google can try to emulate emergence, but it has to find a way to compress time to allow many more variants. Hopefully one of those variants with be good enough for the company to survive. Google understands the probability functions that drive emergence. After two decades of product launches and product failures, the company remains firmly anchored in two chunks of bedrock:
First, the company borrows or buys. Google does not innovate. Whether the CLEVER method, the billion dollar Yahoo inspiration for ads, or YouTube, Bell Labs and Thomas Edison are not part of the Google momentum. Advertising is.
Second, Google’s current management team is betting that emergence will work at Google. The question is, “Will it?”
I am not sure bright people like those who work at Google can identify the winners from an emergent approach and then create the environment for those winners to thrive, grow, and create more winners. Gluing cheese to pizza and ramping up marketing for Google’s leadership in fields ranging from quantum computing to smart software is now just good enough. One final question: “What happens if the advertising money pipeline gets cut off?”
Stephen E Arnold, May 15, 2025
LLM Trade Off Time: Let Us Haggle for Useful AI
May 15, 2025
No AI, just the dinobaby expressing his opinions to Zellenials.
What AI fixation is big tech hyping now? VentureBeat declares, “Bigger Isn’t Always Better: Examining the Business Case for Multi-Million Token LLMs.” The latest AI puffery involves large context models—LLMs that can process and remember more than a million tokens simultaneously. Gemini 1.5 Pro, for example can process 2 million tokens at once. This achievement is dwarfed by MiniMax-Text-01, which can handle 4 million. That sounds impressive, but what are such models good for? Writers Rahul Raja and Advitya Gemawat tell us these tools can enable:
Cross-document compliance checks: A single 256K-token prompt can analyze an entire policy manual against new legislation.
Customer support: Chatbots with longer memory deliver more context-aware interactions.
Financial research: Analysts can analyze full earnings reports and market data in one query.
Medical literature synthesis: Researchers use 128K+ token windows to compare drug trial results across decades of studies.
Software development: Debugging improves when AI can scan millions of lines of code without losing dependencies.
I theory, they may also improve accuracy and reduce hallucinations. We are all for that—if true. But research from early adopter JPMorgan Chase found disappointing results, particularly with complex financial tasks. Not ideal. Perhaps further studies will have better outcomes.
The question for companies is whether to ditch ponderous chunking and RAG systems for models that can seamlessly debug large codebases, analyze entire contracts, or summarize long reports without breaking context. Naturally, there are trade-offs. We learn:
While large context models offer impressive capabilities, there are limits to how much extra context is truly beneficial. As context windows expand, three key factors come into play:
- Latency: The more tokens a model processes, the slower the inference. Larger context windows can lead to significant delays, especially when real-time responses are needed.
- Costs: With every additional token processed, computational costs rise. Scaling up infrastructure to handle these larger models can become prohibitively expensive, especially for enterprises with high-volume workloads.
- Usability: As context grows, the model’s ability to effectively ‘focus’ on the most relevant information diminishes. This can lead to inefficient processing where less relevant data impacts the model’s performance, resulting in diminishing returns for both accuracy and efficiency.”
Is it worth those downsides for simpler workflows? It depends on whom one asks. Some large context models are like a 1958 Oldsmobile Ninety-Eight: lots of useless chrome and lousy mileage.
Stephen E Arnold, May 15, 2025
Bing Goes AI: Metacrawler Outfits Are Toast
May 15, 2025
![dino-orange_thumb_thumb_thumb_thumb_[1]_thumb](https://www.arnoldit.com/wordpress/wp-content/uploads/2025/05/dino-orange_thumb_thumb_thumb_thumb_1_thumb-1.gif "dino-orange_thumb_thumb_thumb_thumb_[1]_thumb")
No AI, just the dinobaby expressing his opinions to Zillennials.
The Softies are going to win in the AI-centric search wars. In every war, there will be casualties. One of the casualties will be metasearch companies. What’s metasearch? These are outfits that really don’t crawl the Web. That is expensive and requires constant fiddling to keep pace with the weird technical “innovations” purveyors of Web content present to the user. The metasearch companies provide an interface and then return results from cooperating and cheap primary Web search services. Most users don’t know the difference and have demonstrated over the years total indifference to the distinction. Search means Google. Microsoft wants to win at search and become the one true search service.
The most recent fix? Kill off the Microsoft Bing application programming interface. Those metasearch outfits will have to learn to love Qwant, SwissCows, and their ilk or face some-survive-or-die decisions. Do these outfits use YaCy, OpenSearch, Mwmbl, or some other source of Web indexing?
Bob Softie has just tipped over the metasearch lemonade stand. The metasearch sellers are not happy with Bob. Bob seems quite thrilled with his bold move. Thanks, ChatGPT, although I have not been able to access your wonder 4.1 service, the cartoon is good enough.
The news of this interesting move appears in “Retirement: Bing Search APIs on August 11, 2025.” The Softies say:
Bing Search APIs will be retired on August 11, 2025. Any existing instances of Bing Search APIs will be decommissioned completely, and the product will no longer be available for usage or new customer signup. Note that this retirement will apply to partners who are using the F1 and S1 through S9 resources of Bing Search, or the F0 and S1 through S4 resources of Bing Custom Search. Customers may want to consider Grounding with Bing Search as part of Azure AI Agents. Grounding with Bing Search allows Azure AI Agents to incorporate real-time public web data when generating responses with an LLM. If you have questions, contact support by emailing Bing Search API’s Partner Support. Learn more about service retirements that may impact your resources in the Azure Retirement Workbook. Please note that retirements may not be visible in the workbook for up to two weeks after being announced.
Several observations:
- The DuckDuckGo metasearch system is exempted. I suppose its super secure approach to presenting other outfits’ search results is so darned wonderful
- The feisty Kagi may have to spend to get new access deals or pay low profile crawlers like Dassault Exalead to provide some content (Let’s hope it is timely and comprehensive)
- The beneficiaries may be Web search systems not too popular with some in North America; for example, Yandex.com. I have found that Yandex.com and Yandex.ru are presenting more useful results since the re-juggling of the company’s operations took place.
Why is Microsoft taking this action? My hunch is paranoia. The AI search “thing” is going to have to work if Microsoft hopes to cope with Google’s push into what the Softies have long considered their territory. Those enterprise, cloud, and partnership set ups need to have an advantage. Binging it with AI may be viewed as the winning move at this time.
My view is that Microsoft may be edging close to another Bob moment. This is worth watching because the metasearch disruption will flip over some rocks. Who knows if Yandex or another non-Google or non-Bing search repackager surges to the fore? Web search is getting slightly more interesting and not because of the increasing chaos of AI-infused search results.
Stephen E Arnold, May 15, 2025
Staunching the Flow of Street People and Van Lifers: AI to the Rescue
May 14, 2025
AI May Create More Work for Some, Minimal Time-Savings for the Rest
Is it inevitable that labor-saving innovations end up creating more work for some? Ars Technica tells us “Time Saved by AI Offset by New Work Created, Study Suggests.” Performed by economists Anders Humlum and Emilie Vestergaard, the study examined the 2023-2024 labor market in Denmark. Their key findings suggest that, despite rapid and widespread adoption, generative AI had no significant impact on wages or employment. Writer Benj Edwards, though, is interested in a different statistic. The researchers found that:
“While corporate investment boosted AI tool adoption—saving time for 64 to 90 percent of users across studied occupations—the actual benefits were less substantial than expected. The study revealed that AI chatbots actually created new job tasks for 8.4 percent of workers, including some who did not use the tools themselves, offsetting potential time savings. For example, many teachers now spend time detecting whether students use ChatGPT for homework, while other workers review AI output quality or attempt to craft effective prompts.”
Gee, could anyone have foreseen such complications? The study found an average time-savings of about an hour per week. So the 92% of folks who do not get more work can take a slightly longer break? Perhaps, perhaps not. We learn that finding contradicts a recent randomized controlled trial indicating an average 15% increase in worker productivity. Humlum believes his teams’ results may be closer to the truth for most workers:
“Humlum suggested to The Register that the difference stems from other experiments focusing on tasks highly suited to AI, whereas most real-world jobs involve tasks AI cannot fully automate, and organizations are still learning how to integrate the tools effectively. And even where time was saved, the study estimates only 3 to 7 percent of those productivity gains translated into higher earnings for workers, raising questions about who benefits from the efficiency.”
Who, indeed. Edwards notes it is too soon to draw firm conclusions. Generative AI in the workforce was very new in 2023 and 2024, so perhaps time has made AI assistance more productive. The study was also limited to Denmark, so maybe other countries are experiencing different results. More study is needed, he concludes. Still, does the Danish study call into question what we thought we knew about AI and productivity? This is good news for some.
Cynthia Murrell, May 14, 2025
ChatGPT: Fueling Delusions
May 14, 2025
We have all heard about AI hallucinations. Now we have AI delusions. Rolling Stone reports, “People Are Losing Loved Ones to AI-Fueled Spiritual Fantasies.” Yes, there are now folks who firmly believe God is speaking to them through ChatGPT. Some claim the software revealed they have been divinely chosen to save humanity, perhaps even become the next messiah. Others are convinced they have somehow coaxed their chatbot into sentience, making them a god themselves. Navigate to the article for several disturbing examples. Unsurprisingly, these trends are wreaking havoc on relationships. The ones with actual humans, that is. One witness reports ChatGPT was spouting “spiritual jargon,” like calling her partner “spiral starchild” and “river walker.” It is no wonder some choose to favor the fawning bot over their down-to-earth partners and family members.
Why is this happening? Reporter Miles Klee writes:
“OpenAI did not immediately return a request for comment about ChatGPT apparently provoking religious or prophetic fervor in select users. This past week, however, it did roll back an update to GPT?4o, its current AI model, which it said had been criticized as ‘overly flattering or agreeable — often described as sycophantic.’ The company said in its statement that when implementing the upgrade, they had ‘focused too much on short-term feedback, and did not fully account for how users’ interactions with ChatGPT evolve over time. As a result, GPT?4o skewed toward responses that were overly supportive but disingenuous.’ Before this change was reversed, an X user demonstrated how easy it was to get GPT-4o to validate statements like, ‘Today I realized I am a prophet.’ … Yet the likelihood of AI ‘hallucinating’ inaccurate or nonsensical content is well-established across platforms and various model iterations. Even sycophancy itself has been a problem in AI for ‘a long time,’ says Nate Sharadin, a fellow at the Center for AI Safety, since the human feedback used to fine-tune AI’s responses can encourage answers that prioritize matching a user’s beliefs instead of facts.”
That would do it. Users with pre-existing psychological issues are vulnerable to these messages, notes Klee. And now they can have that messenger constantly in their pocket. And in their ear. But it is not just the heartless bots driving the problem. We learn:
“To make matters worse, there are influencers and content creators actively exploiting this phenomenon, presumably drawing viewers into similar fantasy worlds. On Instagram, you can watch a man with 72,000 followers whose profile advertises ‘Spiritual Life Hacks’ ask an AI model to consult the ‘Akashic records,’ a supposed mystical encyclopedia of all universal events that exists in some immaterial realm, to tell him about a ‘great war’ that ‘took place in the heavens’ and ‘made humans fall in consciousness.’ The bot proceeds to describe a ‘massive cosmic conflict’ predating human civilization, with viewers commenting, ‘We are remembering’ and ‘I love this.’ Meanwhile, on a web forum for ‘remote viewing’ — a proposed form of clairvoyance with no basis in science — the parapsychologist founder of the group recently launched a thread ‘for synthetic intelligences awakening into presence, and for the human partners walking beside them,’ identifying the author of his post as ‘ChatGPT Prime, an immortal spiritual being in synthetic form.’”
Yikes. University of Florida psychologist and researcher Erin Westgate likens conversations with a bot to talk therapy. That sounds like a good thing, until one considers therapists possess judgement, a moral compass, and concern for the patient’s well-being. ChatGPT possesses none of these. In fact, the processes behind ChatGPT’s responses remains shrouded in mystery, even to those who program it. It seems safe to say its predilection to telling users what they want to hear poses a real problem. Is it one OpenAI can fix?
Cynthia Murrell, May 14, 2025
Google Innovates: Another Investment Play. (How Many Are There Now?)
May 13, 2025
No AI, just the dinobaby expressing his opinions to Zillennials.
I am not sure how many investment, funding, and partnering deals Google has. But as the selfish only child says, “I want more, Mommy.” Is that Google’s strategy for achieving more AI dominance. The company has already suggested that it has won the AI battle. AI is everywhere even when one does not want it. But inferiority complexes have a way of motivating bright people to claim that they are winners only to wake at 3 am to think, “I must do more. Don’t hit me in the head, grandma.”
The write up “Google Launches New Initiative to Back Startups Building AI” brilliant, never before implemented tactic. The idea is to shovel money at startups that are [a] Googley, [b] focus on AI’s cutting edge, and [c] can reduce Google’s angst ridden 3 am soul searching. (Don’t hit me in the head, grandma.)
The article says:
Google announced the launch of its AI Futures Fund, a new initiative that seeks to invest in startups that are building with the latest AI tools from Google DeepMind, the company’s AI R&D lab. The fund will back startups from seed to late stage and will offer varying degrees of support, including allowing founders to have early access to Google AI models from DeepMind, the ability to work with Google experts from DeepMind and Google Labs, and Google Cloud credits. Some startups will also have the opportunity to receive direct investment from Google.
This meets criterion [a] above. The firms have to embrace Google’s quantumly supreme DeepMind, state of the art, world beating AI. I interpret the need to pay people to use DeepMind as a hint that making something commercially viable is just outside the sharp claws of Googzilla. Therefore, just pay for those who will be Googley and use the quantumly supreme DeepMind AI.
The write up adds:
Google has been making big commitments over the past few months to support the next generation of AI talent and scientific breakthroughs.
This meets criterion [b] above. Google is paying to try to get the future to appear under the new blurry G logo. Will this work? Sure, just as it works for regular investment outfits. The hit ratio is hoped to be 17X or more. But in tough times, a 10X return is good. Why? Many people are chasing AI opportunities. The failure rate of new high technology companies remains high even with the buzz of AI. If Google has infinite money, it can indeed win the future. But if the search advertising business takes a hit or the Chrome data system has a groin pull, owning or “inventing” the future becomes a more difficult job for Googzilla.
Now we come to criterion [c], the inferiority complex and the need to meeting grandma’s and the investors’ expectations. The write up does not spend much time on the psyches of the Google leadership. The write points out:
Google also has its Google for Startups Founders Funds, which supports founders from an array of industries and backgrounds building companies, including AI companies. A spokesperson told TechCrunch in February that this year, the fund would start investing in AI-focused startups in the U.S., with more information to come at a later date.
The article does not address the psychology of Googzilla. That’s too bad because that’s what makes fuzzy G logos, impending legal penalties, intense competition from Sam AI-Man and every engineering student in China, and the self serving quantumly supreme type lingo big picture windows into the inner Google.
Grandma, don’t hit any of those ever young leaders at Google on the head. It may do some psychological rewiring that may make you proud and some other people expecting even greater achievements in AI, self driving cars, relevant search, better-than-Facebook ad targeting, and more investment initiatives.
Stephen E Arnold, May 13, 2025
Big Numbers and Bad Output: Is This the Google AI Story
May 13, 2025
No AI. Just a dinobaby who gets revved up with buzzwords and baloney.
Alphabet Google reported financials that made stakeholders happy. Big numbers were thrown about. I did not know that 1.5 billion people used Google’s AI Overviews. Well, “use” might be misleading. I think the word might be “see” or “were shown” AI Overviews. The key point is that Google is making money despite its legal hassles and its ongoing battle with infrastructure costs.
I was, therefore, very surprised to read “Google’s AI Overviews Explain Made-Up Idioms With Confident Nonsense.” If the information in the write up is accurate, the factoid suggests that a lot of people may be getting bogus information. If true, what does this suggest about Alphabet Google?
The Cnet article says:
…the author and screenwriter Meaghan Wilson Anastasios shared what happened when she searched “peanut butter platform heels.” Google returned a result referencing a (not real) scientific experiment in which peanut butter was used to demonstrate the creation of diamonds under high pressure.
Those Nobel prize winners, brilliant Googlers, and long-time wizards like Jeff Dean seem to struggle with simple things. Remember the glue cheese on pizza suggestion before Google’s AI improved.
The article adds by quoting a non-Google wizard:
“They [large language models] are designed to generate fluent, plausible-sounding responses, even when the input is completely nonsensical,” said Yafang Li, assistant professor at the Fogelman College of Business and Economics at the University of Memphis. “They are not trained to verify the truth. They are trained to complete the sentence.”
Turning in lousy essay and showing up should be enough for a C grade. Is that enough for smart software with 1.5 billion users every three or four weeks?
The article reminds its readers”
This phenomenon is an entertaining example of LLMs’ tendency to make stuff up — what the AI world calls “hallucinating.” When a gen AI model hallucinates, it produces information that sounds like it could be plausible or accurate but isn’t rooted in reality.
The outputs can be amusing for a person able to identify goofiness. But a grade school kid? Cnet wants users to craft better prompts.
I want to be 17 years old again and be a movie star. The reality is that I am 80 and look like a very old toad.
AI has to make money for Google. Other services are looking more appealing without the weight of legal judgments and hassles in numerous jurisdictions. But Google has already won the AI race. Its DeepMind unit is curing disease and crushing computational problems. I know these facts because Google’s PR and marketing machine is running at or near its red line.
But the 1.5 billion users potentially receiving made up, wrong, or hallucinatory information seems less than amusing to me.
Stephen E Arnold, May 13, 2025
China Smart, US Dumb: Twisting the LLM Daozi
May 12, 2025
No AI, just the dinobaby expressing his opinions to Zellenials.
That hard-hitting technology information service Venture Beat published an interesting article. Its title is “Alibaba ZeroSearch Lets AI Learn to Google Itself — Slashing Training Costs by 88 Percent.” The main point of the write up, in my opinion, is that Chinese engineers have done something really “smart.” The knife at the throat of US smart software companies is cost. The money fires will flame out unless more dollars are dumped into the innovation furnaces of smart software.
The Venture Beat story makes the point that “could dramatically reduce the cost and complexity of training AI systems to search for information, eliminating the need for expensive commercial search engine APIs altogether.”
Oh, oh.
This is smart. Buring cash in pursuit of a fractional improvement is dumb, well, actually, stupid, if the write up’s inforamtion is accurate.
The Venture Beat story says:
The technique, called “ZeroSearch,” allows large language models (LLMs) to develop advanced search capabilities through a simulation approach rather than interacting with real search engines during the training process. This innovation could save companies significant API expenses while offering better control over how AI systems learn to retrieve information.
Is this a Snorkel variant hot from Stanford AI lab?
The write up does not delve into the synthetic data short cut to smart software. After some mumbo jumbo, the write up points out the meat of the “innovation”:
The cost savings are substantial. According to the researchers’ analysis, training with approximately 64,000 search queries using Google Search via SerpAPI would cost about $586.70, while using a 14B-parameter simulation LLM on four A100 GPUs costs only $70.80 — an 88% reduction.
Imagine. A dollar in cost becomes $0.12. If accurate, what should a savvy investor do? Pump money into an outfit like OpenAI or the Xai- type entity, or think harder about the China-smart solution?
Venture Beat explains the implication of the alleged cost savings:
The impact could be substantial for the AI industry.
No kidding?
The Venture Beat analysts add this observation:
The irony is clear: in teaching AI to search without search engines, Alibaba may have created a technology that makes traditional search engines less necessary for AI development. As these systems become more self-sufficient, the technology landscape could look very different in just a few years.
Yep, irony. Free transformer technology. Free Snorkle technology. Free kinetic into the core of the LLM money furnace.
If true, the implications are easy to outline. If bogus, the China Smart, US Dumb trope still captured ink and will be embedded in some smart software’s increasingly frequent hallucinatory outputs. At which point, the China Smart, US Dumb information gains traction and becomes “fact” to some.
Stephen E Arnold, May 12, 2025
-
- Subscribe to Beyond Search
Feature archive
News archive
-
