Qwen: Better, Faster, Cheaper. Sure, All Three
September 17, 2025
No smart software involved. Just a dinobaby’s work.
I spotted another China Smart, U S Dumb write up. Analytics India published “Alibaba Introduces Qwen3-Next as a More Efficient LLM Architecture.” The story caught my attention because it was a high five to the China-linked Alibaba outfit and because it is a signal that India and China are on the path to BFF bliss.
The write up says:
Alibaba’s Qwen team has introduced Qwen3-Next, a new large language model architecture designed to improve efficiency in both training and inference for ultra-long context and large-parameter settings.
The sentence reinforces the better, faster, cheaper sales mantra one beloved by Crazy Eddie.
Here’s another sentence catching my attention:
At its core, Qwen3-Next combines a hybrid attention mechanism with a highly sparse mixture-of-experts (MoE) design, activating just three billion of its 80 billion parameters during inference. The announcement blog explains that the new mechanism allows the base model to match, and in some cases outperform, the dense Qwen3-32B, while using less than 10% of its training compute. In inference, throughput surpasses 10x at context lengths beyond 32,000 tokens.
This passage emphasizes the value of the mixture of experts approach in the faster and cheaper assertions.
Do I believe the data?
Sure, I believe every factoid presented in the better, faster, cheaper marketing of large language models. Personally I find that these models, regardless of development group, are useful for some specific functions. The hallucination issue is the deal breaker. Who wants to kill a person because a smart medical system is making benign out of malignancy? Who wants an autonomous AI underwater drone to take out those college students and not the adversary’s stealth surveillance boat?
Where can you get access this better, faster, cheaper winner? The write up says, “Hugging Face, ModelScope, Alibaba Cloud Model Studio and NVIDIA API Catalog, with support from inference frameworks like SGLang and vLLM.”
Stephen E Arnold, September 17, 2025
Professor Goes Against the AI Flow
September 17, 2025
One thing has Cornell professor Kate Manne dreading the upcoming school year: AI. On her Substack, “More to Hate,” the academic insists, “Yes, It Is Our Job as Professors to Stop our Students Using ChatGPT.” Good luck with that.
Manne knows even her students who genuinely love to learn may give in to temptation when faced with an unrelenting academic schedule. She cites the observations of sociologist Tressie McMillan Cottom as she asserts young, stressed-out students should not bear that burden. The responsibility belongs, she says, to her and her colleagues. How? For one thing, she plans to devote precious class time to having students hand-write essays. See the write-up for her other ideas. It will not be easy, she admits, but it is important. After all, writing assignments are about developing one’s thought processes, not the finished product. Turning to ChatGPT circumvents the important part. And it is sneaky. She writes:
“Again, McMillan Cottom crystallized this perfectly in the aforementioned conversation: learning is relational, and ChatGPT fools you into thinking that you have a relationship with the software. You ask it a question, and it answers; you ask it to summarize a text, and it offers to draft an essay; you request it respond to a prompt, using increasingly sophisticated constraints, and it spits out a response that can feel like your own achievement. But it’s a fake relationship, and a fake achievement, and a faulty simulacrum of learning. It’s not going to office hours, and having a meeting of the minds with your professor; it’s not asking a peer to help you work through a problem set, and realizing that if you do it this way it makes sense after all; it’s not consulting a librarian and having them help you find a resource you didn’t know you needed yet. Your mind does not come away more stimulated or enriched or nourished by the endeavor. You yourself are not forging new connections; and it makes a demonstrable difference to what we’ve come to call ‘learning outcomes.’”
Is it even possible to keep harried students from handing in AI-generated work? Manne knows she is embarking on an uphill battle. But to her, it is a fight worth having. Saddle up, Donna Quixote.
Cynthia Murrell, September 17, 2025
Who Needs Middle Managers? AI Outfits. MBAs Rejoice
September 16, 2025
No smart software involved. Just a dinobaby’s work.
I enjoy learning about new management trends. In most cases, these hip approaches to reaching a goal using people are better than old Saturday Night Live skits with John Belushi dressed as a bee. Here’s a good one if you enjoy the blindingly obvious insights of modern management thinkers.
Navigate to “Middle Managers Are Essential for AI Success.” That’s a title for you!
The write up reports without a trace of SNL snarkiness:
31% of employees say they’re actively working against their company’s AI initiatives. Middle managers can bridge the gap.
Whoa, Nellie. I thought companies were pushing forward with AI because, AI is everywhere. Microsoft Word, Google “search” (I use the term as a reminder that relevance is long gone), and from cloud providers like Salesforce.com. (Yeah, I know Salesforce is working hard to get the AI thing to go, and it is doing what big companies like to do: Cut costs by terminating humanoids.)
But the guts of the modern management method is a list (possibly assisted by AI?) The article explains without a bit of tongue in cheek élan “ways managers can turn anxious employees into AI champions.”
Here’s the list:
- Communicate the AI vision. [My observation: Isn’t that what AI is supposed to deliver? Fewer employees, no health care costs, no retirement costs, and no excess personnel because AI is so darned effective?”]
- Say, “I understand” and “Let’s talk about it.” [My observation: How long does psychological- and attitudinal-centric interactions take when there are fires to put out about an unhappy really big customer’s complaint about your firm’s product or service?]
- Explain to the employee how AI will pay off for the employee who fears AI won’t work or will cost the person his/her job? [My observation: A middle manager can definitely talk around, rationalize, and lie to make the person’s fear go away. Then the middle manager will write up the issue and forward it to HR or a superior. We don’t need a weak person on our team, right?]
- “Walk the talk.” [My observation: That’s a variant of fake it until you make it. The modern middle manager will use AI, probably realize that an AI system can output a good enough response so the “walk the talk” person can do the “walk the walk” to the parking lot to drive home after being replaced by an AI agent.]
- Give employees training and a test. [My observation: Adults love going to online training sessions and filling in the on-screen form to capture trainee responses. Get the answers wrong, and there is an automated agent pounding emails to the failing employee to report to security, turn in his/her badge, and get escorted out of the building.]
These five modern management tips or insights are LinkedIn-grade output. Who will be the first to implement these at an AI company or a firm working hard to AI-ify its operations. Millions I would wager.
Stephen E Arnold, September 16, 2025
Google Is Going to Race Penske in Court!
September 15, 2025
Written by an unteachable dinobaby. Live with it.
How has smart software affected the Google? On the surface, we have the Code Red klaxons. Google presents big time financial results so the sirens drowned out by the cheers for big bucks. We have Google dodging problems with the Android and Chrome snares, so the sounds are like little chicks peeping in the eventide.
—-
FYI: The Penske Outfits
- Penske Corporation itself focuses on transportation, truck leasing, automotive retail, logistics, and motorsports.
- Penske Media Corporation (PMC), a separate entity led by Jay Penske, owns major media brands like Rolling Stone and Billboard.
—-
What’s actually going on is different, if the information in “Rolling Stone Publisher Sues Google Over AI Overview Summaries.” [Editor’s note: I live the over over lingo, don’t you?] The write up states:
Google has insisted that its AI-generated search result overviews and summaries have not actually hurt traffic for publishers. The publishers disagree, and at least one is willing to go to court to prove the harm they claim Google has caused. Penske Media Corporation, the parent company of Rolling Stone and The Hollywood Reporter, sued Google on Friday over allegations that the search giant has used its work without permission to generate summaries and ultimately reduced traffic to its publications.
Site traffic metrics are an interesting discipline. What exactly are the log files counting? Automated pings, clicks, views, downloads, etc.? Google is the big gun in traffic, and it has legions of SEO people who are more like cheerleaders for making sites Googley, doing the things that Google wants, and pitching Google advertising to get sort of reliable traffic to a Web site.
The SEO crowd is busy inventing new types of SEO. Now one wants one’s weaponized content to turn up as a link, snippet, or footnote in an AI output. Heck, some outfits are pitching to put ads on the AI output page because money is the name of the game. Pay enough and the snippet or summary of the answer to the user’s prompt may contain a pitch for that item of clothing or electronic gadget one really wants to acquire. Psychographic ad matching is marvelous.
The write up points out that an outfit I thought was into auto racing and truck rentals but is now a triple threat in publishing has a different take on the traffic referral game. The write up says:
Penske claims that in recent years, Google has basically given publishers no choice but to give up access to its content. The lawsuit claims that Google now only indexes a website, making it available to appear in search, if the publisher agrees to give Google permission to use that content for other purposes, like its AI summaries. If you think you lose traffic by not getting clickthroughs on Google, just imagine how bad it would be to not appear at all.
Google takes a different position, probably baffled why a race car outfit is grousing. The write up reports:
A spokesperson for Google, unsurprisingly, said that the company doesn’t agree with the claims. “With AI Overviews, people find Search more helpful and use it more, creating new opportunities for content to be discovered. We will defend against these meritless claims.” Google Spokesperson Jose Castaneda told Reuters.
Gizmodo, the source for the cited article about the truck rental outfit, has done some original research into traffic. I quote from the cited article:
Just for kicks, if you ask Google Gemini if Google’s AI Overviews are resulting in less traffic for publishers, it says, “Yes, Google’s AI Overview in search results appears to be resulting in less traffic for many websites and publishers. While Google has stated that AI Overviews create new opportunities for content discovery, several studies and anecdotal reports from publishers suggest a negative impact on traffic.”
I have some views on this situation, and I herewith present them to you:
- Google is calm on the outside but in crazy mode internally. The Googlers are trying to figure out how to keep revenues growing as referral traffic and the online advertising are undergoing some modest change. Is the glacier calving? Yep, but it is modest because a glacier is big and the calf is small.
- The SEO intermediaries at the Google are communicating like Chatty Cathies to the SEO innovators. The result will be a series of shotgun marriages among the lucrative ménage à trois of Google’s ad machine, search engine optimization professional, and advertising services firms in order to lure advertisers to a special private island.
- The bean counters at Google are looking at their MBA course materials, exam notes for CPAs, and reading books about forensic accounting in order to make the money furnaces at Google hot using less cash as fuel. This, gentle reader, is a very, very difficult task. At another time, a government agency might be curious about the financial engineering methods, but at this time, attention is directed elsewhere I presume.
Net net: This is a troublesome point. Google has lots of lawyers and probably more cash to spend on fighting the race car outfit and its news publications. Did you know that the race outfit owned the definitive publication about heavy metal as well at Billboard magazine?
Stephen E Arnold, September 15, 2025
Shame, Stress, and Longer Hours: AI’s Gifts to the Corporate Worker
September 15, 2025
Office workers from the executive suites to entry-level positions have a new reason to feel bad about themselves. Fortune reports, “ ‘AI Shame’ Is Running Rampant in the Corporate Sector—and C-Suite Leaders Are Most Worried About Getting Caught, Survey Says.” Writer Nick Lichtenberg cites a survey of over 1,000 workers by SAP subsidiary WalkMe. We learn almost half (48.8%) of the respondents said they hide their use of AI at work to avoid judgement. The number was higher at 53.4% for those at the top—even though they use AI most often. But what about the generation that has entered the job force amid AI hype? We learn:
“Gen Z approaches AI with both enthusiasm and anxiousness. A striking 62.6% have completed work using AI but pretended it was all their own effort—the highest rate among any generation. More than half (55.4%) have feigned understanding of AI in meetings. … But only 6.8% report receiving extensive, time-consuming AI training, and 13.5% received none at all. This is the lowest of any age group.”
In fact, the study found, only 3.7% of entry-level workers received substantial AI training, compared to 17.1% of C-suite executives. The write-up continues:
“Despite this, an overwhelming 89.2% [of Gen Z workers] use AI at work—and just as many (89.2%) use tools that weren’t provided or sanctioned by their employer. Only 7.5% reported receiving extensive training with AI tools.”
So younger employees use AI more but receive less training. And, apparently, are receiving little guidance on how and whether to use these tools in their work. What could go wrong?
From executives to fresh hires and those in between, the survey suggests everyone is feeling the impact of AI in the workplace. Lichtenberg writes:
“AI is changing work, and the survey suggests not always for the better. Most employees (80%) say AI has improved their productivity, but 59% confess to spending more time wrestling with AI tools than if they’d just done the work themselves. Gen Z again leads the struggle, with 65.3% saying AI slows them down (the highest amount of any group), and 68% feeling pressure to produce more work because of it.”
In addition, more than half the respondents said AI training initiatives amounted to a second, stressful job. But doesn’t all that hard work pay off? Um, no. At least, not according to this report from MIT that found 95% of AI pilot programs at large companies fail. So why are we doing this again? Ask the investor class.
Cynthia Murrell, September 15, 2025
How Much Is That AI in the Window? A Lot
September 15, 2025
AI technology is expensive. Big Tech companies are aware of the rising costs, but the average organization is unaware of how much AI will make their budgets skyrocket. The Kilo Code blog shares insights into AI’s soaring costs in, “Future AI Bills Of $100K/YR Per Dev.”
Kilo recently broke the 1 trillion tokens a month barrier on OpenRouter for the first time. Other open source AI coding tools experienced serious growth too. Claude and Cursor “throttled” their users and encouraged them to use open source tools. These AI algorithms needed to be throttled because their developers didn’t anticipate that application inference costs would rise. Why did this happen?
“Application inference costs increased for two reasons: the frontier model costs per token stayed constant and the token consumption per application grew a lot. We’ll first dive into the reasons for the constant token price for frontier models and end with explaining the token consumption per application. The price per token for the frontier model stayed constant because of the increasing size of models and more test-time scaling. Test time scaling, also called long thinking, is the third way to scale AI…While the pre- and post-training scaling influenced only the training costs of models. But this test-time scaling increases the cost of inference. Thinking models like OpenAI’s o1 series allocate massive computational effort during inference itself. These models can require over 100x compute for challenging queries compared to traditional single-pass inference.”
If organizations don’t want to be hit with expensive AI costs they should consider using open source models. Open source models ere designed to assist users instead of throttling them on the back send. That doesn’t even account for people expenses such as salaries and training.
Costs and customers’ willingness to pay escalating and unpredictable fees for AI may be a problem the the AI wizards cannot explain away. Those free and heavily discounted deals may deflate some AI balloons.
Whitney Grace, September 15, 2025
China Smart, US Dumb: The Baidu AI Service
September 12, 2025
It seems smart software is good for something. CNBC reports, “AI Avatars in China Just Proved They Are Ace Influencers: It Only Took a Duo 7 Hours to Rake in More than $7 Million.” Chinese tech firm Baidu collaborated with two human influencers on the project. Reporter Evelyn Cheng tells us:
“Luo Yonghao, one of China’s earliest and most popular live streamers, and his co-host Xiao Mu both used digital versions of themselves to interact with viewers in real time for well over six hours on Sunday on Baidu’s e-commerce livestreaming platform ‘Youxuan’, the Chinese tech company said. The session raked in 55 million yuan ($7.65 million). In comparison, Luo’s first livestream attempt on Youxuan last month, which lasted just over four hours, saw fewer orders for consumer electronics, food and other key products, Baidu said.”
The experiment highlights Baidu’s avatar technology, which can save marketing departments a lot of money. We learn:
“Luo’s and his co-host’s avatars were built using Baidu’s generative AI model, which learned from five years’ worth of videos to mimic their jokes and style, Wu Jialu, head of research at Luo’s other company, Be Friends Holding, told CNBC on Wednesday. … AI avatars can sharply reduce costs since companies don’t need to hire a large production team or a studio to livestream. The digital avatars can also stream nonstop without needing breaks. … [Wu] said that Baidu now offers the best digital human product currently available, compared to the early days of livestreaming e-commerce five or six years ago.”
Yes, the “early” days of five or six years ago, when the pandemic forced companies and workers to explore their online options. Both landed on livestreaming to generate sales and commissions. Now, it seems, companies can cut the human talent out of the equation. How efficient.
Cynthia Murrell, September 12, 2025
AI Algorithms Are Not Pirates, Just Misunderstood
September 11, 2025
Let’s be clear: AI algorithms are computer programs designed to imitate human brains. They’re not sentient . They are taught using huge amounts of data sets that contain pirated information. By proxy this makes AI developers thieves. David Carson on Medium wrote, “Theft Is Not Fair Use” arguing that AI is not abiding by one of the biggest laws that powers YouTube. (One of the big AI outfits just wrote a big check for unauthorized content suck downs. Not guilty, of course.)
Publishers, record labels, entertainment companies, and countless artists are putting AI developers on notice by filing lawsuits against AI developers. Thomson Reuters was victorious against an AI-based legal platform, Ross Intelligence, for harvesting its data. It’s a drop in the water bucket however, because Trump’s Artificial Intelligence Action Plan sought input from Big Tech. Open AI and Google asked to be exempt from copyright in their big data sets. A group of authors are suing Meta and a copyright law professor gaggle filed an amicus brief on their behalf. The professors poke holes in Meta’s fair use claim.
Big Tech is powerful and they’ve done this for years:
"Tech companies have a history of taking advantage of legacy news organizations that are desperate for revenue and are making deals with short-term cash infusions but little long-term benefit. I fear AI companies will act as vampires, draining news organizations of their valuable content to train their new AI models and then ride off into the sunset with their multi-billion dollar valuations while the news organizations continue to teeter on the brink of bankruptcy. It wouldn’t be the first time tech companies out-maneuvered (online advertising) or lied to news organizations.”
Unfortunately creative types are probably screwed. What’s funny is that Carson is a John S. Knight Journalism Fellow at Stanford. It’s the same school in which the president manipulated content to advance his career. How many of these deep suckers are graduates of this esteemed institution? Who teaches copyright basics? Maybe an AI system?
Whitney Grace, September 11, 2025
AI a Security Risk? No Way or Is It No WAI?
September 11, 2025
Am I the only one who realizes that AI is a security problem? Okay, I’m not but organizations certainly aren’t taking AI security breaches says Venture Beat in the article, “Shadow AI Adds $670K To Breach Costs While 97% Of Enterprises Skip Basic Access Controls, IBM Reports.” IBM collected information with the Ponemon Institute (does anyone else read that as Pokémon Institute?) about data breaches related to AI. IBM and the Ponemon Institute held 3470 interviews with 600 organizations that had data breaches.
Shadow AI is the unauthorized use of AI tools and applications. IBM shared how shadow AI affects organizations in the Cost of a Data Breach Report. Unauthorized usage of AI tools cost organizations $4.63 million and that is 16% more than the $4.44 million global average. YIKES! Another frightening statistic is that 97% of the organizations lacked proper AI access controls. Only 13% had AI-security related breaches compared to 8% who were unaware if AI comprised their systems
Bad actors are using supply chains as their primary attack and AI allows them to automate tasks to blend in with regular traffic. If you want to stay awake at night here are some more numbers:
“A majority of breached organizations (63%) either don’t have an AI governance policy or are still developing one. Even when they have a policy, less than half have an approval process for AI deployments, and 62% lack proper access controls on AI systems.”
An expert said this about the issue:
This pattern of delayed response to known vulnerabilities extends beyond AI governance to fundamental security practices. Chris Goettl, VP Product Management for Endpoint Security at Ivanti, emphasizes the shift in perspective: ‘What we currently call ‘patch management’ should more aptly be named exposure management—or how long is your organization willing to be exposed to a specific vulnerability?’”
Organizations that are aware of AI breaches and have security plans in place save more money.
It pays to be prepared and cheaper too!
Whitney Grace, September 11, 2025
Microsoft: The Secure Discount King
September 10, 2025
Just a dinobaby sharing observations. No AI involved. My apologies to those who rely on it for their wisdom, knowledge, and insights.
Let’s assume that this story in The Register is dead accurate. Let’s forget that Google slapped the $0.47 smart software price tag on its Gemini smart software. Now let’s look at the interesting information in “Microsoft Rewarded for Security Failures with Another US Government Contract.” Snappy title. But check out the sub-title for the article: “Free Copilot for Any Agency Who Actually Wants It.”
I did not know that a US government agency was human signaled by the “who.” But let’s push forward.
The article states:
The General Services Administration (GSA) announced its new deal with Microsoft on Tuesday, describing it as a “strategic partnership” that could save the federal government as much as $3.1 billion over the next year. The GSA didn’t mention specific discount terms, but it said that services, including Microsoft 365, Azure cloud services, Dynamics 365, Entra ID Governance, and Microsoft Sentinel, will be cheaper than ever for feds. That, and Microsoft’s next-gen Clippy, also known as Copilot, is free to access for any agency with a G5 contract as part of the new deal, too. That free price undercuts Google’s previously cheapest-in-show deal to inject Gemini into government agencies for just $0.47 for a year.
Will anyone formulate the hypothesis that Microsoft and Google are providing deep discounts to get government deals and the every-popular scope changes, engineering services, and specialized consulting fees?
I would not.
I quite like comparing Microsoft’s increasingly difficult to explain OpenAI, acqui-hire, and home-grown smart software as Clippy. I think that the more apt comparison is the outstanding Microsoft Bob solution to interface complexity.
The article explains that Oracle landed contracts with a discount, then Google, and now Microsoft. What about the smaller firms? Yeah, there are standard procurement guidelines for those outfits. Follow the rules and stop suggesting that giant companies are discounting there way into the US government.
What happens if these solutions hallucinate, do not deliver what an Inspector General, an Independent Verification & Validation team, or the General Accounting Office expects? Here’s the answer:
With the exception of AWS, all the other OneGov deals that have been announced so far have a very short shelf life, with most expirations at the end of 2026. Critics of the OneGov program have raised concerns that OneGov deals have set government agencies up for a new era of vendor lock-in not seen since the early cloud days, where one-year discounts leave agencies dependent on services that could suddenly become considerably more expensive by the end of next year.
The write up quotes one smaller outfit’s senior manager’s concern about low prices. But the deals are done, and the work on the 2026-2027 statements of work has begun, folks. Small outfits often lack the luxury of staff dedicated to extending a service provider’s engagement into a year or two renewal target.
The write up concludes by bringing up ancient history like those pop archaeologists on YouTube who explain that ancient technology created urns with handles. The write up says:
It was mere days ago that we reported on the Pentagon’s decision to formally bar Microsoft from using China-based engineers to support sensitive cloud services deployed by the Defense Department, a practice Defense Secretary Pete Hegseth called “mind-blowing” in a statement last week. Then there was last year’s episodes that allowed Chinese and Russian cyber spies to break into Exchange accounts used by high-level federal officials and steal a whole bunch of emails and other information. That incident, and plenty more before it, led former senior White House cyber policy director AJ Grotto to conclude that Microsoft was an honest-to-goodness national security threat. None of that has mattered much, as the feds seem content to continue paying Microsoft for its services, despite wagging their finger at Redmond for “avoidable errors.”
Ancient history or aliens? I don’t know. But Microsoft does deals, and it is tough to resist “free”.
Stephen E Arnold, September 10, 2025
-
- Subscribe to Beyond Search
Feature archive
News archive
-
