Tech Brew keeps business leaders up-to-date on the latest innovations, automation advances, policy shifts, and more, so they can make informed decisions about tech.

Tech Brew

Patrick Kulp

techbrew.com

Accenture’s chief AI officer thinks every C-suite needs a job like hers

Lan Guan is leading the consultancy’s efforts to harness a huge wave of AI interest from the business world.

A video of a sailboat moving down a river generated with Runway's Gen-2

Video generated with Runway's Gen-2 of a "person running on a track."

Video generated from the prompt "cat running on the street."

Is video the next frontier for generative AI?

AI startups want to make inroads in Hollywood, but the technology still has its limits.

Persistent AI hype fails to buoy startup market in Q1

PitchBook data shows a sluggish market for venture capital investing and fundraising.

For the last few months, Walmart has been parading out an army of new agents for its workers, according to global chief technology and development officer Suresh Kumar. AI systems can now help with tasks like pricing items, stocking shelves, or requesting time off.

But the whole ensemble may have gotten to be a bit much.

“If you build individual agents for every use case, it very quickly becomes overwhelming for the end user,” Kumar told us.

Now, a team of “super agents” is here to save the day. The retail giant is reorganizing its many agents under four of these souped-up personas, which can route queries to any of the dozens of sub-agents under their scope.

The four super agents will be geared toward shoppers, store associates, suppliers and advertisers, and developers. Already live is “

,” a shopping assistant that can help plan around an occasion, summarize reviews, or set up recurring household orders. “Marty” will help merchants craft ad campaigns and list items. The two internal super agents, for employees and developers, remain nameless.

“The agent knows that you are a merchant. The agent knows that you are a store manager. The agent knows that you work in people services. And it also knows all our sub-agents that are there. It can coordinate,” Kumar said. “We will have tons and tons of sub-agents underneath as we start automating more capabilities, as we start addressing more use cases.”

 reached a dull roar in the tech industry, experts have been 

 this tech is all about teamwork. Initial agentic experiments will eventually level up into orchestrated networks of task-specific entities, these pros said.

“The dominant discourse in modern LLM-based agents is still leaning heavily toward building large single agents as one-stop shops, and I don’t think that scales,” Babak Hodjat, CTO of AI at Cognizant, told us in a previous interview in February. “Multi-agent systems hold the key to leveraging AI in ways that can balance and analyze often conflicting needs to deliver higher value impact.”

 But most companies are far from that point. A recent EY survey found that only 14% of companies have fully implemented agents. The biggest barriers are all trust-related, EY found—cybersecurity and data privacy concerns chief among them.

Walmart has a framework to govern who has access to what when it comes to security and compliance, Kumar said. In most cases, there’s still a human in the loop to evaluate response quality and fine-tune the model based on where it might be “misaligned,” Kumar said.

“[We] realize that this is a journey, just like all of these other businesses,” Kumar said. “So it’s not like we are going to start on day one giving full access and full delegate authority and full capabilities to any single agent, even the super agent.”

Walmart is building its agents to coordinate with data sources through Anthropic’s Model Context Protocol (MCP), an open-source framework that standardizes how agents connect with outside systems. Kumar said the plan is to make the system fully interoperable so that, for example, Sparky might continue a shopping journey that started on another site, or the associates’ super agent might communicate with an Aetna AI agent about health insurance benefits.

As the shopping giant continues to build out these agentic systems, Kumar envisions the shopping experience eventually encompassing different modes of media, from TikTok videos to snapped photos of home decor.

“If I’ve already decided on a particular decor for my living room, what is the particular painting that I want to hang on the wall? This is not something that you can…do through search keywords,” he said. “No, you actually take a picture and you let the agent go, ‘Ah, now I understand what it is that you are trying to do: Here, let me suggest a bunch of stuff.’ So I think that is the future.”

Innumerable AI agents can overwhelm. Can Walmart’s ‘super agents’ come to the rescue?

The shopping giant is reorganizing its AI capabilities under four “super agents.”

 are a natural use case for coding. But how do developers actually feel about the technology that’s shaking up their jobs?

Stack Overflow asked 49,000 of them in its annual survey, and found some mixed opinions. Adoption was indeed growing; around 84% used or planned to use AI in their development process, compared to 76% last year. But positive sentiments were in decline (70% → 60% favorability), while mistrust in its accuracy was on the rise (31% distrusted it in 2024, whereas 46% were skeptical this year.)

Only about 31% of developers said they used agents at least once a month, but around 38% said they had no plans to start doing so. A little over half of those who use them said agents have had a positive effect on their productivity.

“We know more developers are using AI tools, but they’re not having exactly a great experience using those tools,” Erin Yepis, senior analyst of market research and insights at Stack Overflow, told Tech Brew. “There’s more developers that have the ability to try it, and maybe their expectations aren’t exactly where the reality lies.”

 As far as what’s holding back agentic adoption, developers identified accuracy and security or privacy concerns as their biggest problems with these more autonomous systems.

“One of the top frustrations that developers have is when the AI code isn’t quite right, they have to spend more time sort of debugging the thing that’s supposed to help them save time and get work done faster,” Yepis said.

 about declining employment in the tech industry, most developers (64%) said AI is not a threat to their jobs, though that number decreased slightly from 68% last year. Yepis said the survey might be showing that developers are becoming a bit more hesitant about the way AI coding tools are changing their workflows.

“There are just a lot more tools out there now, and I think [they’re] becoming a little bit more wary, as far as, ‘Maybe part of my job is going to change—what part of the job? What is it going to look like in three to five years?” Yepis said.

Developers responding to the survey seemed to overwhelmingly reject the term “

” to describe what they do; 72% said the loosely defined practice is not part of their development work. The term became popular earlier this year to describe a certain improvisational approach to AI coding, but the trend has also 

. For the Stack Overflow survey, the vibe coding question offered open-ended responses and drew many strong opinions, Yepis said.

“Whether it’s just that the term itself is off-putting or not, I thought that was interesting,” Yepis said. “There’s passionate responses when it comes to developers and their jobs.”

A set of hands on a keyboard with AI code overlaid

Most developers aren’t using agents, but many who do find them useful

But a Stack Overflow survey also found that positive sentiment toward the tech has dipped.

When it comes to AI, Big Tech has to spend money to make money (eventually…fingers crossed).

Companies like Microsoft, Google, and Meta have gotten that first part down, pouring 

 into new data centers and other AI infrastructure. Microsoft and Google saw those bets start to pay off nicely in the most recent quarter, with major growth in their respective cloud businesses. Meta also claimed its strong ad sales were fueled by AI efficiencies.

Investors cheered strong results from Meta and Microsoft this week with more than $500 billion in stock market gains, according to 

 Microsoft’s Azure cloud business pulled in $75 billion in the past year, notching 34% YoY growth. Along with strong overall earnings and revenue beats, the results fueled the tech giant’s 

 as the second-ever $4 trillion company this week, behind Nvidia. This was the first time Microsoft broke out revenue for Azure specifically, beyond the “cloud services” bucket.

Microsoft has said it would slow down capital expenditures in the next fiscal year, which started this month, but it still plans to spend $30 billion this quarter. CFO Amy Hood said the company is trying to keep up with a backlog of existing demand.

Microsoft said its Copilot apps now have 100 million users every month.

 While Microsoft plans to slow spending, Google said last week that its AI infrastructure spending would be $85 billion for the year, $10 billion more than previously expected, because of strong cloud demand. Google also marked a banner quarter for its cloud arm with 32% YoY growth, as well as broader earnings and revenue beats.

Alphabet recently notched a win when OpenAI said it would add Google Cloud to its roster of suppliers for ChatGPT, as OpenAI continues to diversify beyond its partnership with Microsoft, 

 Then there’s Meta, where strong earnings and revenue beats ensured that Mark Zuckerberg was able to spend most of the call discussing his favorite topic: the company’s new quest for superintelligence. Meta has spent recent weeks setting its game board with 

 organized around former Scale AI CEO Alexandr Wang.

All those massive pay packages for AI researchers will make employee compensation Meta’s second biggest cost driver for next year, beyond AI infrastructure.

In a missive Zuckerberg shared ahead of earnings, the Meta CEO said “novel safety concerns” will mean the company has to be “careful about what we choose to open source.” Meta’s Llama model family has been among the leading options for 

He clarified in the call that while “I don’t think that our thinking has particularly changed” on open-source, some models might be too big or unsafe to open-source in the future. “We’re getting models that are so big that they’re just not practical for a lot of other people to use. So it’s, we would kind of wrestle with whether it’s productive or helpful to share that or if that’s really just primarily helping competitors or something like that,” Zuckerberg said.

Google, Meta, Microsoft on 3 pillars surrounded by binary code

The cloud business is booming for Microsoft and Google

During an earnings call, Mark Zuckerberg revealed more about Meta’s quest for superintelligence.

Can an AI chatbot act as your therapist? 93 million bucks says the answer is yes.

That’s how much money Andreessen Horowitz (a16z), Radical Ventures, Forerunner Ventures, and other firms have poured into Slingshot AI, a mental health startup from Casper co-founder Neil Parikh and AI researcher Daniel Cahn. The company just released its first AI chatbot tool, dubbed Ash, after 18 months of development with 50,000 beta users.

Parikh told Tech Brew that Slingshot was born from a desire to help ease a shortage of mental health providers. He hopes Ash might eventually expand access to certain mental health support to those with issues like depression, anxiety, or social isolation, freeing human therapists to focus on patients with the most serious needs.

 But how can a technology that’s reportedly 

 also provide mental health assistance? One key, according to Parikh, is to equip the AI with the ability to push back on its users.

; their built-in yes-man nature might inadvertently egg people down paths of delusion or paranoia. Parikh said Slingshot’s team has taken pains to ensure that Ash will challenge its users when the situation calls for it.

And unlike ChatGPT, Ash, which is available through text or voice, also ends conversational sessions after it deems a certain amount of progress is made, according to Parikh. “We don’t want to build your best friend or your companion,” he said. “The goal is for Ash to actually help you build relationships in the real world.”

 Daisy Wolf, an a16z investment partner specializing in healthcare technology, said she sees mental health AI as a major investment opportunity, given the shortage of professionals in the space.

“The demand is massive, but we’ll never have enough trained professionals to meet it in our lifetimes. The only way to scale support meaningfully is through technology,” Wolf said in an email. “We believe generative AI has the potential to radically expand access to mental health support…but only if it’s built with deep clinical rigor and consumer trust at the core.”

Bryan Kim, an a16z partner on the consumer AI side, said he sees the dataset that Slingshot has assembled as its competitive advantage.

“These guys are training the model, and they’re obtaining the data; that’s really hard,” Kim said. “That’s a really expensive and difficult path to go on, versus ‘Oh, here’s an LLM—let’s post-train it a little bit; let’s fork it a little bit.’”

 Ash consists of a foundation model fine-tuned on therapy data in “dozens of therapeutic styles and approaches” provided by partner organizations. That dataset is about 1% of the size of the base model’s, which is usually at least trillions of tokens—the equivalent of short words or parts of words. Slingshot also has a clinical team that writes custom responses to help train the model.

The model also uses reinforcement learning with human feedback (RLHF) to continually improve based on user interactions.

 Parikh pointed out that the quest to build a therapybot is far from new, dating back at least to a 

 at MIT in the 1960s called ELIZA. It’s also 

startup currently attempting mental health assistance with AI.

Recent studies have started to evaluate the usefulness of these tools (not including Ash), and the results have been mixed. A 

 warned that AI therapy chatbots tend to be ineffective and dangerous. But a New England Journal of Medicine 

 found the technology “promising” for certain contexts, like anxiety, depression, and eating disorders.

As Ash rolls out, Parikh said Slingshot will be collecting feedback from hundreds of customers every day about how it’s working and what to do differently. It’s free now, but the company plans to eventually monetize it.

“We just want to hear from people about what’s valuable to them and what they want out of a product like this,” he said.

Image of transparent brains blending into each other.

Slingshot AI's Ash wants to be your therapist

Slingshot AI is betting that ChatGPT-like tools can perform mental health services—with the right guardrails.

This well-funded chatbot wants to be your therapist

Here at Tech Brew, our editorial process heavily features the AI transcription software Otter, through which we run recordings of our interviews. It’s not perfect, by any means—our account has an inexplicable preoccupation with mishearing words as “Covid,” for one (more on that later)—but it definitely beats the arduous process of transcribing by hand.

As tech companies seek to make generative AI commonplace in everyday home and work life, speech recognition is a sizable part of that mission, whether it’s 

. And ever since deep learning breakthroughs in the 2010s made tools like Otter widespread and then transformers improved them further, this capability has been taken somewhat for granted—it’s maybe not the flashiest subfield of AI in 2025.

But George Saon, a distinguished research scientist who leads IBM’s speech strategy, told Tech Brew it’s a mistake to treat speech recognition as a solved problem. Issues like discerning overlapping voices, making out different accents, and dealing with background noise can still make real-world use cases dicey, he said.

“It’s interesting, because every few years, someone claims that speech recognition has been solved, that we’re at the point where we’re just as good as humans, so we have to move on,” Saon told us. “And at IBM, we don’t believe that. We think there’s still a ways to go to get there. And the reason is that speech is complex, because it has many dimensions of variability.”

At the time of this conversation, IBM’s latest Granite speech recognition model was the leading system on Hugging Face’s leaderboard (it’s now in second place). Saon said his team has zeroed in on a number of improvements that have boosted their models’ accuracy, from adding more robust training data across different dialects to using LLMs to suss out the context of a given conversation.

 In 2021, IBM and McDonald’s announced a partnership to test automated drive-thru ordering in certain restaurants. After a nearly three-year test, the fast food chain 

Saon said the task of taking orders was fraught with challenges. There was the background noise: McDonald’s locations might be near noisy airports or train stations; fire trucks and ambulances could screech by. Sometimes, drive-thrus had two side-by-side booths and orders would be confused between them. The system could also mistakenly pick up on side chatter within the cars.

Different dialects and accents make understanding more difficult as well. “Our models now are multi-dialect,” Saon said. “They can work just as well for Australian English, Indian, UK—whatever accent you want—that works now, but that wasn’t the case awhile ago.”

It’s generally harder to parse speech than written text, Saon said. “Then you have, um, people, when they, they talk—they—it’s not like written text, right? It’s—they’re often agrammatical, they, they, they have repetitions, deletions, hesitations. It’s, it’s, it’s messy,” Saon said. (This quote was transcribed exactly, and it’s what most of our quotes would look like without the light editing out of verbal tics we do for nearly everyone.)

 IBM is exploring how LLMs can help voice recognition systems better understand the context behind a given audio clip, Saon said. For instance, IBM’s model could guess the correct spelling of the golfer Jack Nicklaus’s name if it ascertains that a conversation is about golf. (For the record, Otter’s transcription got the spelling right half the time.)

One surprising revelation was that integrating multilingual LLMs allowed the model to interpret code-switching—in this sense, shifting between languages mid-sentence or phrase—without any additional training, Saon said.

In order to prime the model for some of those problems it might encounter in the real world, Saon said the team will introduce noises, distortions, and gaps into the training data. He said there’s no algorithmic workaround for simply adding more training data in the different ways that people speak or audio is captured.

“We take training utterances, and we artificially add various types of noises from babble speech or cars, or all sorts of things, to make the model more robust to noisy conditions,” Saon said. “So data variety…and perturbation, augmentation of the acoustic signal.”

Improving this accuracy is becoming more important as generative AI becomes 

—able to field requests by voice or visually in addition to text. While text is still the dominant way that people interact with LLMs, companies want to add voices to AI for everything from smart speakers to customer service reps—and drive-thru orders. Contextual understanding could unlock new gains in making these systems more precise, Saon said.

“Context is everything,” he said. “Contextual [automatic speech recognition] is going to be a big thing in the future.”

As for Otter’s penchant for mistakenly hearing “Covid” with disproportionate frequency, we reached out to the company to try to understand the AI’s behavior. Otter spokesperson Mitchell Woodrow took the issue to Otter’s engineering team and came back with a likely explanation: Otter may be fixated on Covid simply because 

 have also been for much of the past half-decade—and the model learns from us.

“In cases of unclear audio, [automatic speech recognition] models may default to highly recognizable terms like ‘Covid,’ Woodrow said. “The term ‘Covid’ became one of the most frequently spoken words globally in recent years, which means it's statistically more likely to appear when the model is uncertain. Our system is designed to continuously adapt, which allowed us to implement a quick fix in this case.”

Image of a mouth speaking with an error message to represent use of AI in speech recognition technology. 

Why IBM says speech recognition is not a fully solved problem

And what the company learned from McDonald’s drive-thru orders.

Since nobody is exactly holding their breath for a big AI safety push in Congress these days, observers agree that the path to AI regulation in the US 

. That path looked perilously close to being blocked in early July, until the Senate overwhelmingly 

 a federal budget bill moratorium that would have blocked state and local AI rules for a decade.

With that proposed ban dead—for now—experts say a surge of state-level laws may be on the way. That’s in spite of some 

 against enforcing such regulations in Trump’s 

. The focus of state legislation has so far ranged from targeted rules around pressing use cases like deepfakes and AI in healthcare or hiring, to comprehensive bills that cover high-risk AI models.

 Hope Anderson, a partner in White & Case’s data, privacy, and cybersecurity practice, compared the push to state-level efforts to legislate digital privacy in lieu of congressional action. AI-focused legislation touches on a wider range of areas, however, which is already leading to more variety in emerging state laws, she said.

“States are accelerating their efforts to regulate AI, and I expect that trend to continue, likely outpacing what we saw with privacy,” Anderson told Tech Brew in an email. “Unlike privacy laws, which tend to follow a more uniform structure, AI touches such a wide array of sectors that, in addition to comprehensive AI laws, we’re already seeing legislation emerge on everything from automated decision-making to deepfakes and healthcare. The breadth of use cases likely means we’re going to see a surge in diverse, state-specific laws.”

 Forrester principal analyst Alla Valente said to expect states to begin “fast-tracking stalled bills or by taking up new ones” in the wake of the moratorium’s failure.

 that piecemeal regulation will lead to a maze of different rules for companies to navigate state by state—it was part of the justification for the proposed moratorium in the first place. Valente said this situation might be avoided if states take cues from one another’s landmark laws.

“To avoid a federal patchwork of AI laws, we could see interstate coordination where bills such as the New York RAISE Act or existing laws in California or Colorado become the template for other states,” she said.

Last fall, California Governor Gavin Newsom vetoed what would have been a 

. That bill’s author, State Senator Scott Wiener, is back with a narrower bill that focuses on whistleblower rights for AI employees flagging dangers, and advancing a public cloud computing cluster. In an emailed statement to Tech Brew, Wiener criticized Trump’s AI Action Plan for “bullying states into not regulating AI.”

New York is on the verge of enacting a safety bill called the RAISE Act, which would be the first law to impose legal safety standards on frontier AI models. The bill focuses specifically on catastrophic scenarios involving mass casualties or billion-dollar damages. It requires developers of models that cost more than $100 million to publish certain safety protocols and contingency plans as well as annual reviews. The bill has 

 the state legislature and is awaiting Governor Kathy Hochul’s signature.

What to expect from state AI bills in wake of the failed moratorium

Experts say states could accelerate efforts to weigh in on key risks.

By the end of this year, you might walk into a theater and see a big-budget movie that was made in some significant part with generative AI. You may not even notice; like early CGI, the technology’s fingerprints will only draw the eye when it’s done badly.

These are the predictions of Naeem Talukdar, co-founder and CEO of Moonvalley, a startup that makes “commercially safe” AI models for Hollywood studios and other creative businesses. Nearly every studio is already tapping AI for early-stage previsualization work, he claimed. But copyright concerns have stopped moviemakers from going all-in on the technology in actual production.

“The biggest blocker to date for studios has been the legal piece,” Talukdar said. “It’s that their legal teams just can’t sign off on it.”

That could change with the rise of video generation tools that purport to be “clean,” or trained only on licensed data, like Moonvalley’s new Marey model or 

 to train AI on copyrighted material, these models let risk-averse businesses dodge iffy IP questions.

 Moonvalley recently raised $84 million, days after the public release of its Marey video model. Talukdar said the company trained Marey entirely on videos licensed from the IP holder for the length of training, usually around a couple of years (the license to generate with the resulting model is perpetual). It also doesn’t ingest public domain works or its user-generated content, Talukdar said.

The data pool comes from a range of sources: independent studios, filmmakers, and videographers; film communities across the world; “film students who have tons of B-roll…that they’re not doing anything with,” Talukdar said.

“People aren’t really worried about accidentally generat[ing], like, 

, or something like that, because you’ll find that and you’ll remove it,” Talukdar said. “The real concern, especially for these studios, is…what if it’s just some random thing nobody can recognize that was copyrighted 30 or 40 years ago? Basically, this risk that I might just have infringing content all over the place, and I have no idea how to control for that.”

Lydia Clougherty Jones, a senior director analyst at research firm Gartner, said businesses of all kinds remain very concerned about legal questions around AI. Even using ‘non-copyrightable’ material can be tricky, she said.

“Risk of copyright infringement continues to be top of mind,” Clougherty Jones said in an email. “A risk-averse perspective would be to use only non-copyrightable material, but even that approach still requires assurances that the content is indeed not ‘copyrightable,’ which can be challenging given the nuances that distinguish what is and what is not copyrightable.”

 Copyright law is, of course, far from the only fraught topic around AI in moviemaking, which has raised issues from 

 and aesthetic concerns to questions about the nature of human creativity. When Asteria, Moonvalley’s partner AI studio, announced Natasha Lyonne’s directorial debut, she faced 

“It’s obviously a sensitive topic, as it should be,” Talukdar said. “You look at the previous [entertainment guild] strikes, you look at everything that’s kind of come around it. And I also think a big part of why we exist is that tech companies have been very irresponsible with how they roll out these models.”

While Asteria describes Moonvalley’s models as “ethical,” Talukdar said he’d rather not weigh in on the ethics of AI in film.

“We prefer to use the words like ‘commercially safe,’” Talukdar said. “The ethical side? My opinion on that is that my opinion shouldn’t matter, in that the less tech people opine on that, probably the better. I think that it’s up to the industry to figure out what’s right, and it’s up to us to abide by that.”

That said, he’s “100% optimistic” that AI will be a good thing for the entertainment industry, adding that AI will make “the areas of production that require the least sort of creative input” cheaper, like establishing shots, B-roll, and backgrounds. He’s convinced good acting and writing will remain impossible to replace.

“It’s not this idea that the big studios are now just gonna start factory-farming content slop and printing profits. It’s also not the flip side, which is, every Joe like me is going to suddenly be creating movies left and right,” Talukdar said. “What’s really going to happen is what we’re already seeing start to happen, is this middle layer of people who have talent, who have taste, who have creativity, but they don’t have necessarily that access, who don’t have a connection to one of the big five studios in the world, will now have the ability to create things that they couldn’t create before.”

3D movie glasses with binary code in the lenses

AI