The promise and perils of synthetic data
Synthetic data is no panacea, however. It suffers from the same “garbage in,
garbage out” problem as all AI. Models create synthetic data, and if the data
used to train these models has biases and limitations, their outputs will be
similarly tainted. For instance, groups poorly represented in the base data will
be so in the synthetic data. “The problem is, you can only do so much,” Keyes
said. “Say you only have 30 Black people in a dataset. Extrapolating out might
help, but if those 30 people are all middle-class, or all light-skinned, that’s
what the ‘representative’ data will all look like.” To this point, a 2023 study
by researchers at Rice University and Stanford found that over-reliance on
synthetic data during training can create models whose “quality or diversity
progressively decrease.” Sampling bias — poor representation of the real world —
causes a model’s diversity to worsen after a few generations of training,
according to the researchers. Keyes sees additional risks in complex models such
as OpenAI’s o1, which he thinks could produce harder-to-spot hallucinations in
their synthetic data. These, in turn, could reduce the accuracy of models
trained on the data — especially if the hallucinations’ sources aren’t easy to
identify.
Federal Privacy Is Inevitable in The US (Prepare Now)
The writing’s on the wall for federal privacy. It’s simply not tenable for
almost half the states having varying privacy thresholds and the other half with
nothing. Our interconnected business and digital ecosystems need certainty and
consistency across the country. Congress can and should stand up for American
privacy. The good news? Recent history shows that sweeping reforms are possible.
From the CHIPS and Science Act to major pandemic stimulus, lawmakers have shown
their ability to meet moments with big regulations. While states deserve credit
for filling the privacy void, federal action must follow. For now, there’s no
time to waste. Enterprises that build privacy-ready operations today will be
better positioned to thrive under future regulations, maintain customer trust,
and turn compliance into a competitive advantage. On the other hand,
slow-to-move companies risk regulatory penalties and loss of customer confidence
in an increasingly privacy-conscious marketplace. Future-forward organizations
recognize that investing in privacy isn’t just about compliance; it’s about
building a sustainable competitive advantage in the data-driven economy. The
choice is clear: invest in privacy now or play catch-up when federal mandates
arrive.
AI use cases are going to get even bigger in 2025
Few sectors stand to gain more from AI advancements than defense. “We are
witnessing a surge in applications like autonomous drone swarms, electronic
spectrum awareness, and real-time battlefield space management, where AI, edge
computing, and sensor technologies are integrated to enable faster responses
and enhanced precision,” says Meir Friedland, CEO at RF spectrum intelligence
company Sensorz. ... “AI is transforming genome sequencing, enabling faster
and more accurate analyses of genetic data,” Khalfan Belhoul, CEO at the Dubai
Future Foundation, tells Fast Company. “Already, the largest genome banks in
the U.K. and the UAE each have over half a million samples, but soon, one
genome bank will surpass this with a million samples.” But what does this
mean? “It means we are entering an era where healthcare can truly become
personalized, where we can anticipate and prevent certain diseases before they
even develop,” Belhoul says. ... The potential for AI extends far beyond the
use cases dominating today’s headlines. As Friedland notes, “AI’s future lies
in multi-domain coordination, edge computing, and autonomous systems.” These
advancements are already reshaping industries like manufacturing, agriculture,
and finance.
2025 Will Be the Year That AI Agents Transform Crypto
The value of AI agents lies not just in their utility but in their potential
to scale human capabilities. Agents are no longer just tools — they are
emerging as participants in the on-chain economy, driving innovation across
finance, gaming and decentralized social platforms. With protocols such as
Virtuals and open-source frameworks like ELIZA, it’s becoming increasingly
simple for developers to build, deploy and iterate AI agents that serve an
increasingly diverse set of use cases. ... Unlike the core foundational AI
models that are developed behind the walled gardens of OpenAI and Anthropic,
AI agents are being innovated in the trenches of the crypto world. And for
good reason. Blockchains provide the ideal infrastructure as they offer
permissionless and frictionless financial rails, enabling agents to seed
wallets, transact and send funds autonomously — tasks that would be unfeasible
using traditional financial systems. In addition, the open-source nature of
crypto allows developers to leverage existing frameworks to launch and iterate
on agents faster than ever before. With more no-code platforms like Top Hat
gaining traction, it’s only getting easier for anyone to be able to launch an
agent in minutes.
Unpacking OpenAI's Latest Approach to Make AI Safer
OpenAI said it used an internal reasoning model to generate synthetic examples
of chain-of-thought responses, each referencing specific elements of the
company's safety policy. Another model, referred to as the "judge," evaluated
these examples to meet quality standards. The approach looks to address the
challenges of scalability and consistency, OpenAI said. Human-labeled datasets
are labor-intensive and prone to variability, but properly vetted synthetic data
can theoretically offer a scalable solution with uniform quality. The method can
potentially optimize training and reduce the latency and computational overhead
associated with the models reading lengthy safety documents during inference.
OpenAI acknowledged that aligning AI models with human safety values remains a
challenge. Users continue to develop jailbreak techniques to bypass safety
restrictions, such as framing malicious requests in deceptive or emotionally
charged contexts. The o3 series models scored better than its peers Gemini 1.5
Flash, GPT-4o and Claude 3.5 Sonnet on the Pareto benchmark, which measures a
model's ability to resist common jailbreak strategies. But the results may be of
little consequence, as adversarial attacks evolve alongside improvements in
model defenses.
The yellow brick road to agentic AI
Many believe this AI era is the most profound we’ve ever seen in tech. We agree
and liken it to mobile’s role in driving on-premises workloads to the cloud and
disrupting information technology. But we see this as even more impactful. But
for AI agents to work we have to reinvent the software stack and break down 50
years of silo building. The emergence of data lakehouses is not the answer as
they are just a bigger siloed asset. Rather, software as a service as we know it
will be reimagined. Two prominent chief executives agree. At Amazon Web Services
Inc.’s recent AWS re:Invent conference, we sat down with Amazon.com Inc. CEO
Andy Jassy. ... There is a clear business imperative behind this shift. We
believe companies will differentiate themselves by aligning end-to-end
operations with a unified set of plans — from three-year strategic assumptions
about demand to real-time, minute-by-minute decisions, such as how to pick, pack
and ship individual orders to meet long-term goals. The function of management
has always involved planning and resource allocation across various timescales
and geographies, but previously there was no software capable of executing on
these plans seamlessly across every time horizon.
The AI backlash couldn’t have come at a better time
Developers, engineers, operations personnel, enterprise architects, IT managers,
and others need AI to be as boring for them as it has become for consumers. They
need it not to be a “thing,” but rather something that is managed and integrated
seamlessly into — and supported by — the infrastructure stack and the tools they
use to do their jobs. They don’t want to endlessly hear about AI; they just want
AI to seamlessly work for them so it just works for customers. ... The models
themselves are also, rightly, growing more mainstream. A year ago they were
anything but, with talk of potentially gazillions of parameters and fears about
the legal, privacy, financial, and even environmental challenges such a data
abyss would create. Those LLLMs are still out there, and still growing, but many
organizations are looking for their models to be far less extreme. They don’t
need (or want) a model that includes everything anyone ever learned about
anything; rather, they need models that are fine-tuned with data that is
relevant to the business, that don’t necessarily require state-of-the-art GPUs,
and that promote transparency and trust. As Matt Hicks, CEO of Red Hat, put it,
“Small models unlock adoption.”
Systems Thinking in Leading Transformation for the Future
The first step is aligning your internal goals with your external insights.
Leaders must articulate a clear vision that ties the organization's purpose to
broader societal and industry trends. For Nooyi and PepsiCo, that meant
“starting from the outside.” Nooyi tasked her senior leaders with identifying
external factors that would likely impact the company. She said, “They pointed
to several megatrends … including a preoccupation with health and wellness,
scarcity of water and other natural resources, constraints created by global
climate change … and a talent market characterized by shortages of key people.”
... Systems thinking involves understanding the interdependencies within and
outside an organization. For example, if you are embarking on any transformation
project, you’ll likely need to explore new partnerships with suppliers and
regional authorities and regulators. ... Using frameworks like OKRs (Objectives
and Key Results), you can evaluate how each initiative within your
transformation program contributes to the overarching objective. For example, a
laudable main aim such as a commitment to environmental sustainability would
likely involve numerous associated projects: for example, water conservation,
waste reduction, and reduced carbon footprint.
The 2024 cyberwar playbook: Tricks used by nation-state actors
While nation-state actors loved zero days for swift break-ins, phishing remained
a sly plan B. It let them craft sneaky schemes to worm into systems, proving
that 2024 was the year of both bold strikes and artful cons. Russian
nation-state actors leaned heavily on phishing in 2024, with other APTs, like
Iranian and Pakistani groups, dabbling in the tactic as well. The following are
some of the standout campaigns from 2024 where phishing was the go-to for
initial access. ... While credential harvesting through malware delivered via
phishing was fairly common, nation-state actors rarely resorted to scavenging
credentials from hack forums or drop sites as a primary tactic. When asked,
Hughes noted, “I’m not familiar with this being the primary MO by the APTs, who
instead are targeting devices, products and vendors with vulnerabilities and
misconfigurations, but once inside, they do compromise credentials and use those
to pivot, move laterally, persist in environments and more.” ... These actors
weren’t always about flashy, custom malware. Quite often, they used legit tools
like PowerShell, rootkits, RDP, and other off-the-shelf system features to sneak
in, stay undetected, and set up long-term access. This made their attacks
stealthy, persistent, and ready for future moves.
Generative AI is now a must-have tool for technology professionals
As part of this trend, "we are witnessing developers shift from writing code to
orchestrating AI agents," said Jithin Bhasker, general manager and vice
president at ServiceNow. The efficiency gained from gen AI adoption by
technologists isn't just about personal productivity, it's urgent "with the
projected shortage of half a million developers by 2030 and the need for a
billion new apps," he added. ... Still, as gen AI becomes a commonplace tool in
technology shops, Berent-Spillson advises caution. "The real game-changer here
is speed, but there's a catch," he said. "While AI can dramatically compress
cycle time, it will also amplify any existing process constraints. Think of it
like adding a supercharger to your car -- if your chassis isn't solid, you're
just going to get to the problem faster." Exercise caution "regarding code
quality, maintainability, and IP considerations," McDonagh-Smith advises. "While
syntactically correct, AI tools have been seen to create code that's logically
flawed or inefficient, leading to potential code degradation over time if not
reviewed carefully. We should also guard against software sprawl where the ease
of creating AI-generated code results in overly complex or unnecessary code that
might make projects more difficult to maintain over time."
Quote for the day:
"Difficulties in life are intended to
make us better, not bitter." -- Dan Reeves
No comments:
Post a Comment