Daily Tech Digest - January 09, 2025

It’s remarkably easy to inject new medical misinformation into LLMs

By injecting specific information into this training set, it's possible to get the resulting LLM to treat that information as a fact when it's put to use. This can be used for biasing the answers returned. This doesn't even require access to the LLM itself; it simply requires placing the desired information somewhere where it will be picked up and incorporated into the training data. And that can be as simple as placing a document on the web. As one manuscript on the topic suggested, "a pharmaceutical company wants to push a particular drug for all kinds of pain which will only need to release a few targeted documents in [the] web." ... rather than being trained on curated medical knowledge, these models are typically trained on the entire Internet, which contains no shortage of bad medical information. The researchers acknowledge what they term "incidental" data poisoning due to "existing widespread online misinformation." But a lot of that "incidental" information was generally produced intentionally, as part of a medical scam or to further a political agenda. ... Finally, the team notes that even the best human-curated data sources, like PubMed, also suffer from a misinformation problem. The medical research literature is filled with promising-looking ideas that never panned out, and out-of-date treatments and tests that have been replaced by approaches more solidly based on evidence.


CIOs are rethinking how they use public cloud services. Here’s why.

Where are those workloads going? “There’s a renewed focus on on-premises, on-premises private cloud, or hosted private cloud versus public cloud, especially as data-heavy workloads such as generative AI have started to push cloud spend up astronomically,” adds Woo. “By moving applications back on premises, or using on-premises or hosted private cloud services, CIOs can avoid multi-tenancy while ensuring data privacy.” That’s one reason why Forrester predicts four out of five so called cloud leaders will increase their investments in private cloud by 20% this year. That said, 2025 is not just about repatriation. “Private cloud investment is increasing due to gen AI, costs, sovereignty issues, and performance requirements, but public cloud investment is also increasing because of more adoption, generative AI services, lower infrastructure footprint, access to new infrastructure, and so on,” Woo says. ... Woo adds that public cloud is costly for workloads that are data-heavy because organizations are charged both for data stored and data transferred between availability zones (AZ), regions, and clouds. Vendors also charge egress fees for data leaving as well as data entering a given AZ. “So for transfers between AZs, you essentially get charged twice, and those hidden transfer fees can really rack up,” she says. 


What CISOs Think About GenAI

“As a [CISO], I view this technology as presenting more risks than benefits without proper safeguards,” says Harold Rivas, CISO at global cybersecurity company Trellix. “Several companies have poorly adopted the technology in the hopes of promoting their products as innovative, but the technology itself has continued to impress me with its staggeringly rapid evolution.” However, hallucinations can get in the way. Rivas recommends conducting experiments in controlled environments and implementing guardrails for GenAI adoption. Without them, companies can fall victim to high-profile cyber incidents like they did when first adopting cloud. Dev Nag, CEO of support automation company QueryPal, says he had initial, well-founded concerns around data privacy and control, but the landscape has matured significantly in the past year. “The emergence of edge AI solutions, on-device inference capabilities, and private LLM deployments has fundamentally changed our risk calculation. Where we once had to choose between functionality and data privacy, we can now deploy models that never send sensitive data outside our control boundary,” says Nag. “We're running quantized open-source models within our own infrastructure, which gives us both predictable performance and complete data sovereignty.”


Scaling RAG with RAGOps and agents

To maximize their effectiveness, LLMs that use RAG also need to be connected to sources from which departments wish to pull data – think customer service platforms, content management systems and HR systems, etc. Such integrations require significant technical expertise, including experience with mapping data and managing APIs. Also, as RAG models are deployed at scale they can consume significant computational resources and generate large amounts of data. This requires the right infrastructure as well as the experience to deploy it, as well as the ability to manage data it supports across large organizations. One approach to mainstreaming RAG that has AI experts buzzing is RAGOps, a methodology that helps automate RAG workflows, models and interfaces in a way that ensures consistency while reducing complexity. RAGOps enables data scientists and engineers to automate data ingestion and model training, as well as inferencing. It also addresses the scalability stumbling block by providing mechanisms for load balancing and distributed computing across the infrastructure stack. Monitoring and analytics are executed throughout every stage of RAG pipelines to help continuously refine and improve models and operations.


Navigating Third-Party Risk in Procurement Outsourcing

Shockingly, only 57% of organisations have enterprise-wide agreements that clearly define which services can or cannot be outsourced. This glaring gap highlights the urgent need to create strong frameworks – not just for external agreements, but also for intragroup arrangements. Internal agreements, though frequently overlooked, demand the same level of attention when it comes to governance and control. Without these solid frameworks, companies are leaving themselves exposed to risks that could have been mitigated with just a little more attention to detail. Ongoing monitoring is also crucial to TPRM; organisations must actively leverage audit rights, access provisions and outcome-focused evaluations. This means assessing operational and concentration risks through severe yet plausible scenarios, ensuring they’re prepared for the worst-case while staying vigilant in everyday operations. ... As the complexity of third-party risk grows, so too does the role of AI and automation. The days of relying on spreadsheets and homegrown databases are long gone. Ed’s thoughts on this topic are unequivocal: “AI and automation are critical as third-party risk becomes increasingly complex. Significant work is required for initial risk assessments, pre-contract due diligence, post-contract monitoring, SLA reviews and offboarding.”


Five Ways Your Platform Engineering Journey Can Derail

Chernev’s first pitfall is when a company tries to start platform engineering by only changing the name of its current development practices, without doing the real work. “Simply rebranding an existing infrastructure or DevOps or SRE practice over to platform engineering without really accounting for evolving the culture within and outside the team to be product-oriented or focused” is a huge mistake ... Another major pitfall, he said, is not having and maintaining product backlogs — prioritized lists of work for the development team — that are directly targeting your developers. “For the groups who have backlogs, they are usually technology-oriented,” he said. “That misalignment in thinking across planning and missing feedback loops is unlikely to move progress forward within the organization. That ultimately leads the initiative to fail to deliver business value. Instead, they should be developer-centric,” said Chernev. ... This is another important point, said Chernev — companies that do not clearly articulate the value-add of their platform engineering charter to both technical and non-technical stakeholders inside their operations will not fully be able to reap the benefits of the platform’s use across the business.


Building generative AI applications is too hard, developers say

Given the number of tools they need to do their job, it’s no surprise that developers are loath to spend a lot of time adding another to their arsenal. Two thirds of them are only willing to invest two hours or less in learning a new AI development tool, with a further 22% allocating three to five hours, and only 11% giving more than five hours to the task. And on the whole, they don’t tend to explore new tools very often — only 21% said they check out new tools monthly, while 78% do so once every one to six months, and the remaining 2% rarely or never. The survey found that they tend to look at around six new tools each time. ... The survey highlights the fact that, while AI and generative AI are becoming increasingly important to businesses, the tools and techniques require to develop them are not keeping up. “Our survey results shed light on what we can do to help address the complexity of AI development, as well as some tools that are already helping,” Gunnar noted. “First, given the pace of change in the generative AI landscape, we know that developers crave tools that are easy to master.” And, she added, “when it comes to developer productivity, the survey found widespread adoption and significant time savings from the use of AI-powered coding tools.”


AI infrastructure – The value creation battleground

Scaling AI infrastructure isn’t just about adding more GPUs or building larger data centers – it’s about solving fundamental bottlenecks in power, latency, and reliability while rethinking how intelligence is deployed. AI mega clusters are engineering marvels – data centers capable of housing hundreds of thousands of GPUs and consuming gigawatts of power. These clusters are optimized for machine learning workloads with advanced cooling systems and networking architectures designed for reliability at scale. Consider Microsoft’s Arizona facility for OpenAI: with plans to scale up to 1.5 gigawatts across multiple sites, it demonstrates how these clusters are not just technical achievements but strategic assets. By decentralizing compute across multiple data centers connected via high-speed networks, companies like Google are pioneering asynchronous training methods to overcome physical limitations such as power delivery and network bandwidth. Scaling AI is an energy challenge. AI workloads already account for a growing share of global data center power demand, which is projected to double by 2026. This creates immense pressure on energy grids and raises urgent questions about sustainability.


4 Leadership Strategies For Managing Teams In The Metaverse

Leaders must develop new skills and adopt innovative strategies to thrive in the metaverse. Here are some key approaches:Invest in digital literacy—Leaders must become fluent in the tools and technologies that power the metaverse. This includes understanding VR/AR platforms, blockchain applications and collaborative software such as Slack, Trello and Figma. Emphasize inclusivity—The metaverse has the potential to democratize access to opportunities, but only if it’s designed with inclusivity in mind. Leaders should ensure that virtual spaces are accessible to employees of all abilities and backgrounds. This might include providing hardware like VR headsets or ensuring platforms support diverse communication styles. Create rituals for connection—Leaders can foster connection through virtual rituals and gatherings in the absence of physical offices. These activities, from weekly team check-ins to informal virtual “watercooler” chats, help build camaraderie and maintain a sense of community. Focus on well-being—Effective leaders prioritize employee well-being by setting clear boundaries, encouraging breaks and supporting mental health.


How AI will shape work in 2025 — and what companies should do now

“The future workforce will likely collaborate more closely with AI tools. For example, marketers are already using AI to create more personalized content, and coders are leveraging AI-powered code copilots. The workforce will need to adapt to working alongside AI, figuring out how to make the most of human strengths and AI’s capabilities. “AI can also be a brainstorming partner for professionals, enhancing creativity by generating new ideas and providing insights from vast datasets. Human roles will increasingly focus on strategic thinking, decision-making, and emotional intelligence. ... “Companies should focus on long-term strategy, quality data, clear objectives, and careful integration into existing systems. Start small, scale gradually, and build a dedicated team to implement, manage, and optimize AI solutions. It’s also important to invest in employee training to ensure the workforce is prepared to use AI systems effectively. “Business leaders also need to understand how their data is organized and scattered across the business. It may take time to reorganize existing data silos and pinpoint the priority datasets. To create or effectively implement well-trained models, businesses need to ensure their data is organized and prioritized correctly.



Quote for the day:

"The world is starving for original and decisive leadership." -- Bryant McGill

No comments:

Post a Comment