Artificial intelligence agents are facing significant challenges when tasked with complex responsibilities, as highlighted by a recent study from Langchain. The research indicates that AI performance deteriorates under cognitive overload, with one model's effectiveness dropping to just 2% when managing more than seven domains. This finding emphasizes the need for businesses to design AI systems that can effectively manage complexity rather than assuming that AI can scale to handle human-like multitasking. The implications are particularly relevant for industries reliant on automation, such as customer service and IT operations.
In a related development, researchers from Stanford and the University of Washington have introduced a new AI reasoning model called S1, which can be trained at a fraction of the cost of existing high-end models. This innovation raises concerns about the commoditization of AI, as smaller teams can replicate sophisticated models with minimal resources. Meanwhile, Hugging Face has quickly developed an open-source AI research agent that aims to compete with OpenAI's offerings, showcasing the rapid advancements in AI capabilities and the importance of community contributions in this space.
The podcast also discusses the ongoing legal battles surrounding AI and copyright, particularly the recent ruling in favor of Thomson Reuters against Ross Intelligence. This case underscores the complexities of how AI tools are trained using copyrighted material and sets a precedent that could impact future AI developments. As AI continues to evolve, the legal landscape surrounding its use and the rights of content creators remains a critical area of concern.
Finally, the episode touches on the rising unemployment rate in the IT sector, attributed to the increasing influence of AI and automation. The data reveals a significant jump in unemployment among IT workers, with many routine jobs being automated rather than replaced. This shift highlights the need for IT professionals to upskill and adapt to the changing job market, focusing on areas such as AI integration and cybersecurity, as traditional software development roles decline.
Four things to know today
00:00 AI Agents Struggle Under Pressure—More Complexity Means Less Accuracy
03:23 Big AI, Big Trouble? Low-Cost Models Challenge Industry Leaders
07:17 Apple Faces UK Encryption Fight, CISA’s Role Strengthens, and AI Copyright Battle Heats Up
10:38 AI Disrupts IT Hiring: Fewer Software Jobs, More Automation, and Higher Unemployment
Supported by: https://www.huntress.com/mspradio/
Event: https://nerdiocon.com/
💼 All Our Sponsors
Support the vendors who support the show:
👉 https://businessof.tech/sponsors/
🚀 Join Business of Tech Plus
Get exclusive access to investigative reports, vendor analysis, leadership briefings, and more.
👉 https://businessof.tech/plus
🎧 Subscribe to the Business of Tech
Want the show on your favorite podcast app or prefer the written versions of each story?
📲 https://www.businessof.tech/subscribe
📰 Story Links & Sources
Looking for the links from today’s stories?
Every episode script — with full source links — is posted at:
🎙 Want to Be a Guest?
Pitch your story or appear on Business of Tech: Daily 10-Minute IT Services Insights:
💬 https://www.podmatch.com/hostdetailpreview/businessoftech
🔗 Follow Business of Tech
LinkedIn: https://www.linkedin.com/company/28908079
YouTube: https://youtube.com/mspradio
Bluesky: https://bsky.app/profile/businessof.tech
Instagram: https://www.instagram.com/mspradio
TikTok: https://www.tiktok.com/@businessoftech
Facebook: https://www.facebook.com/mspradionews
Hosted by Simplecast, an AdsWizz company. See pcm.adswizz.com for information about our collection and use of personal data for advertising.
[00:00:02] It's Wednesday, February 12th, 2025, and I'm Dave Sobel. Four things to know today. AI agents struggle under pressure. More complexity means less accuracy. Big AI, big trouble. Low-cost models are challenging the industry leaders. Apple faces a UK encryption fight. SIS's role strengthens and AI copyright battle heats up. And
[00:00:25] disrupting AI hiring, fewer software jobs, more automation, and higher unemployment. This is the Business of Tech. A recent study by Langchain reveals that artificial intelligence agents struggle to perform at human levels when overwhelmed by tools and instructions. Findings showed that their performance significantly dropped as agents were tasked with more responsibilities. For instance, one model's
[00:00:51] effectiveness plummeted to just 2% when handling over seven domains. Langchain's tests involved an internal email assistant tasked with responding to customer queries and scheduling meetings, with 30 tasks each run three times. The results indicate that more context leads to worse instruction following, highlighting the challenges of managing multi-agent systems. Langchain aims to explore how to enhance
[00:01:18] agent performance in complex environments. Why do we care? The study confirms that adding more tools and responsibilities doesn't enhance AI performance. It actively degrades it. This aligns with known cognitive overload issues in human workers but presents a more severe drop-off in AI effectiveness. More tools is not the answer. For IT service providers and businesses integrating AI-driven automation,
[00:01:44] it underscores the importance of designing systems that manage complexity effectively rather than assuming AI can simply scale up to human-like multitasking. Now, humans also struggle with overloads, so comparing AI to human multitasking without benchmarking human error rates might be misleading. AI isn't a magic bullet for automation, especially for complex workflows. Investments should focus on structured delegation,
[00:02:12] better context management, and optimized AI agent orchestration rather than assuming a single AI can handle everything. This could shape AI strategy in automation-heavy industries like customer service, IT operations, and enterprise software. With every new breach and threat that I cover, it's clear that cybersecurity isn't a luxury anymore. It's a necessity. That's where Huntress comes in. Their fully
[00:02:39] managed cybersecurity platform is built for every kind of business, not just the 1%. Huntress seamlessly integrates their products and threat hunting team. Their EDR, ITDR, SIM, and security awareness training solutions are purposely built for their elite 24x7 security operations center to stop threats before anyone else even spots them. This potent combination of purpose-built cybersecurity and threat hunting
[00:03:06] expertise is one of the many reasons why G2 users have voted Huntress the number one rated EDR for growing businesses. To see what people-powered cybersecurity looks like, visit Huntress.com slash MSB Radio. A bit more model disruption. Researchers at Stanford and the University of Washington have developed a new
[00:03:30] artificial intelligence reasoning model called S1, which can be trained for under $50 in cloud computing credits. The model demonstrates capabilities similar to high-end models like OpenAI's O1 and DeepSeq's R1, particularly in math and coding tests. The team fine-tuned S1 by distilling it from Google's reasoning model, Gemini 2.0 Flash Thinking Experimental, using a dataset of just 1,000 curated questions.
[00:03:58] NVIDIA H100 GPUs costing around $20 for cloud computing resources. This innovation raises concerns about the commoditization of AI, as small teams can replicate multi-million dollar models at a fraction of the cost. The findings suggest that while distillation is a cost-effective method for recreating existing AI capabilities, it may not lead to the development of significantly superior models. Hugging Face has developed an open source AI research agent named
[00:04:28] OpenDeepResearch in just 24 hours, following the launch of OpenAI's similar feature. This new tool aims to replicate OpenAI's performance, achieving an accuracy of 55% on the general AI assistance benchmark, compared to OpenAI's 76%. The project emphasizes the importance of an agent framework that enables AI models to perform complex, multi-step tasks, such as gathering information from various sources. Hugging Face's Amerik Rocher highlighted that while their current model uses
[00:04:58] closed weights for efficiency, it can easily be adapted to open source models, showcasing the flexibility of their approach. The swift development process was aided by contributions from the open source community. A recent study by the enterprise AI startup Vectara reveals that DeepSeq's R1 model is prone to generating inaccurate information, or hallucinations, at a significantly higher rate than
[00:05:23] other reasoning models. While models from OpenAI and Google showed the least hallucinations, Alibaba's QIN performed the best among those tested. Interestingly, DeepSeq's earlier version, known as V3, was found to be over three times more accurate than R1. Vectara's head of developer relations, Offer Menvolovich, emphasized the importance of balancing various capabilities during model training to mitigate the issues. The findings highlight ongoing concerns regarding DeepSeq's
[00:05:51] training data and moderation capabilities, raising questions about its potential impact on US-based AI companies. Major US cloud providers, including Amazon Web Services, Microsoft and Google, have begun offering access to DeepSeq without charging users based on text generation, unlike other models. Instead, businesses only pay for the computing resources they use. This model can be significantly cheaper, with DeepSeq costing three to four times less than alternatives via an application programming
[00:06:20] interface. For instance, using Meta's Lama model through Amazon Web Services can cost $3 per 1 million tokens, while DeepSeq offers a more efficient processing alternative. Notably, more than 1,000 of Databricks' 12,000 customers have adopted models based on DeepSeq R1. Why do we care? The headline is that the notion that cutting-edge AI requires vast resources is being challenged. This raises existential
[00:06:49] questions for companies investing billions in proprietary models. The models will not be the value in the chain. AI is becoming cheaper and more accessible, but quality and trust remain key differentiators. The industry isn't just competing on raw intelligence anymore, it's competing on efficiency, reliability and cost-effectiveness. And most aren't clearing the bar here. This is all good news for application development and for services providers.
[00:07:18] In a controversial move, the United Kingdom has ordered Apple to create a backdoor that would allow government officials to access encrypted data stored in the cloud by users worldwide. This unprecedented demand issued under the Investigory Powers Act of 2016 requires Apple to provide blanket access to all encrypted content, which critics argue undermines the company's promise of privacy to its users.
[00:07:43] Should Apple comply, it may stop offering encrypted storage in the UK, but this would not satisfy the demand for access in other countries, including the US. Security experts warn that this could create significant vulnerabilities in global cybersecurity. The UK government's insistence on access comes amid growing concerns over encryption being used to shield criminal activities, while tech companies continue to advocate for user privacy and the right to secure communication. The situation escalates fears that
[00:08:12] the UK's action could prompt similar demands from other nations, ultimately endangering user privacy on a global scale. In response to increasing cyber threats, including successful attacks attributed to Chinese hackers, Republicans are softening their criticism of the Cybersecurity and Infrastructure Security Agency, or CISA. Under the new administration, CISA is expected to focus on protecting critical
[00:08:36] infrastructure from hacking attacks. South Dakota Governor Kristi Noem emphasized the agency's essential role in combating ransomware and foreign threats. Despite past calls to dismantle CISA, support for its mission has strengthened, notably from House Homeland Security Committee Chair Mark Green, who highlighted the pressing need to enhance America's electronic defenses amid growing espionage concerns. And Thomson Reuters won a significant early victory in the copyright infringement lawsuit
[00:09:05] against Ross Intelligence, a legal AI startup, marking a pivotal moment in the ongoing legal debates surrounding artificial intelligence and copyright law. The US District Court ruled in favor of Thomson Reuters, stating that Ross's use of its Westlaw search engine content constituted copyright infringement. The judge's decision emphasized that Ross's actions could not be justified under a fair use defense, particularly because they created a direct competitor to Thomson Reuters.
[00:09:34] The case, which has been closely watched as it could set precedence for other similar lawsuits involving major AI firms, highlights the complexity of how AI tools are trained using copyrighted material. Ross, which ceased operations in 2021, argued that its AI was designed to extract legal answers directly from law, but was ultimately found to have significantly copied Thomson Reuters content, which included unique annotations and summaries written by legal experts.
[00:10:03] Why do we care? The UK is a significant market for Apple, so we're waiting to see what they do. In the US, there should be some comfort that SISA may not be as under attack as other agencies. This is all a fast-moving story. And the Thomson Reuters vs. Ross Intelligence case is a big win for traditional copyright holders, setting a precedent that AI models can't indiscriminately scrape and repurpose proprietary content.
[00:10:27] The ruling raises serious questions for AI firms that have trained models on copyrighted datasets. Is generative AI fundamentally at legal risk? JANCO Associates looked at the jobs data I reported earlier this week and reports that the unemployment rate in the information technology sector has risen significantly, climbing from 3.9% in December to 5.7% in January.
[00:10:52] This equates to an increase on unemployed IT workers from 98,000 to 152,000. The rise is attributed to the growing influence of artificial intelligence, which has led to job eliminations in routine tasks, while many companies are opting to automate functions instead of hiring new staff.
[00:11:11] The overall job market added 143,000 jobs during the same period. However, job postings for software development roles have decreased by 8.5% year-over-year. Experts note a bifurcation in job opportunities, where white-collar positions are experiencing less demand compared to in-person, skilled labor jobs.
[00:11:33] Why do we care? A jump from 3.9% to 5.7% in a month is a significant shift. This data is higher than CompTIA's analysis I reported on Monday. Routine IT jobs are being automated rather than replaced, meaning many displaced workers won't find equivalent roles elsewhere. This supports the broader trend of AI streamlining workforces and reducing headcount in traditional IT functions.
[00:11:58] I'll offer that the current slowdown may be a temporary market correction following the overhiring of 2021 to 2022. If AI augmentation rather than full automation becomes the norm, demand for higher-skilled IT professionals may rebound. May. Companies still need AI engineers, cybersecurity experts, and cloud specialists, just not as many traditional software developers.
[00:12:22] For IT professionals, upskilling is now critical, traditional coding jobs are shrinking, and the focus is shifting to AI integration, cybersecurity, and cloud automation. For service providers, expect increased demand for AI-driven efficiency solutions as businesses seek to cut labor costs. However, pricing pressure may intensify as companies reduce IT budgets.
[00:12:48] Are you ready to get your brand in front of the tech leaders shaping the future of managed services? Here at The Business of Tech, we offer flexible sponsorship opportunities to meet your needs, whether it's live show sponsorship, podcast advertising, event promotion, or custom webinars. From affordable exposure options to exclusive sponsorships, our offerings are designed to fit businesses and vendors of all sizes looking to make an impact.
[00:13:15] Prices start at just $500 per month, making our packages a fraction of typical event sponsorship costs. Be a part of the conversation that matters to IT service providers worldwide. Join us at MSP Radio and amplify your message where it counts. Visit mspradio.com slash engage today to explore all the ways we can help you grow. Thanks for listening.
[00:13:44] Today is National Lost Penny Day and National Plum Pudding Day. Nerdy Ocon will be held in Palm Springs, California from April 7th through 9th. Visit nerdyocon.com to learn all about it. The Business of Tech is written and produced by me, Dave Sobel, under ethics guidelines posted at businessof.tech. If you've enjoyed the show, make sure you've subscribed or followed on your favorite platform. It's free and helps directly.
[00:14:14] Give us a review, too. If you want to support the show, visit patreon.com slash mspradio and you'll get access to content early. Or buy our Why Do We Care merch at businessof.tech. Have a question you want answered? We take listener questions, send them in, ideally as a voice memo or video to question at mspradio.com. I answer listener questions live on our Wednesday live show on YouTube and LinkedIn.
[00:14:42] If you've got a comment or a thought on a story, put it in the comments if you're on YouTube or reach out on LinkedIn if you're listening to the podcast. And if you want to advertise on the show, visit mspradio.com slash engage. Once again, thanks for listening and I will talk to you again on our next episode. Part of the MSP Radio Network.

