Latest News

The growing presence of emerging technologies, including artificial intelligence (AI), is raising concerns about their environmental footprint. As states and local governments grapple with the tension between economic development and community health, public health practitioners are well-positioned to help navigate this complex issue.

In many regions, policy responses vary widely, often reflecting a trade-off between furthering state-level economic interests and addressing local community health and environmental concerns. States have primarily focused on energy reporting, ratepayer protection, and environmental assessment, while local governments have tended to act more directly on land use, zoning, and permitting issues.

For example, Loudoun County in Virginia has ended by-right zoning for data centers, requiring all new applications to undergo public hearings starting from 2025. Kansas City in Missouri has classified data centers as industrial, necessitating council approval and mandatory impact studies on water and electricity rates beginning in 2026. Marana in Arizona has prohibited potable water use for cooling and required water source disclosure since 2024.

The path forward for public health practitioners involves leveraging their expertise to address the environmental impacts of emerging technologies. The three core functions of public health – policy development, assessment, and assurance – provide a concrete framework for this work.

In terms of policy development, public health experts can advocate for health impact assessments in permitting processes, transparency requirements, and community notification standards. They should also contribute their expertise to state and local rulemaking efforts.

Assessment is another critical function where public health practitioners can make a significant contribution. This involves tracking and analyzing cumulative environmental exposures, such as air quality near diesel generators, water availability in stressed regions, and electricity cost burdens on low-income households. Public health experts should push for systematic, mandatory data collection from operators to inform these assessments.

Finally, assurance is essential for ensuring that emerging technologies are deployed in a way that prioritizes community health. This involves monitoring health outcomes over time in affected communities, holding operators and regulators accountable to environmental standards, and guaranteeing vulnerable populations have meaningful access to decision-making processes.

The most critical recommendation from public health experts is the need for communities to have the information, access, and standing necessary to participate in decisions about AI infrastructure that will impact their health for decades to come. By serving as a valuable partner in shaping the ethical rollout of emerging technologies, public health can play a crucial role in balancing economic development with community well-being.

AI agent loops have become an essential tool in automating tasks, but choosing the right loop can be a daunting task. With four different types of loops available – turn-based, goal-based, time-based, and proactive – it’s crucial to understand what each one offers and how they differ from one another. In this article, we’ll explore the characteristics of each loop type and provide guidance on selecting the most suitable option for your specific needs.

A new analysis of nearly 10,000 real-world EV battery tests has revealed a surprising trend in the performance of Tesla Model 3 batteries. The study found that the same car model holds up very differently depending on which type of battery it was built with, and the cheaper lithium iron phosphate (LFP) version comes out on top.

The analysis, conducted by Carla, a Swedish used-EV retailer, looked at data from over 9,954 battery tests conducted in Sweden between 2022 and 2026. The tests were done using AVILOO’s battery diagnostics, which measure the actual state of health rather than relying on the car’s dashboard estimate.

When broken down by battery type, the results showed a striking difference in performance. The LFP pack held its charge better than any nickel-based version of the same car, with an average battery health of 93.3% among cars that had driven more than 62,000 miles. This is a significant gap between the best and worst versions of the same car.

The data also revealed that the two nickel-cobalt-aluminum (NCA) packs from Panasonic, which Tesla previously considered its premium option, degraded the most over time. This counterintuitive result challenges the common assumption that more expensive batteries are inherently more durable.

LFP batteries have long been thought to be less durable than their more expensive counterparts due to their lower energy density and higher cost per kWh. However, this study provides a direct apples-to-apples comparison between LFP and nickel-based cells in the same car model.

The key difference between LFP and nickel-based batteries lies in their chemistry. LFP batteries are cheaper and heavier than their nickel-based counterparts but offer better thermal stability and can tolerate full 100% charging without degrading as quickly. This means that owners of Tesla Model 3s with LFP packs may benefit from improved longevity over time.

The finding aligns with previous studies, including a Tesla-funded study and multiple independent teardowns, which have consistently shown that LFP chemistry ages more gradually than nickel-based cells under high mileage conditions. It also suggests that Tesla’s decision to shift its Standard Range Model 3 and Model Y to LFP packs was not just about cost savings but may have provided an added benefit in terms of battery longevity.

The study is part of a broader analysis of the performance of various EV models, with over 20 vehicles included. The Kia e-Niro and Hyundai Kona mechanical twins topped the ranking at over 97% average battery health among cars past 62,000 miles. Every model in the top 20 averaged above 91%, indicating that many modern EVs are capable of retaining a significant portion of their original capacity even after extensive use.

The results also track with other datasets, including Geotab’s telematics study, which found average annual degradation had improved to about 1.8% per year. This suggests that EV batteries may be more durable than previously thought and could potentially last for over two decades or more under normal driving conditions.

A previously unreported Internet-of-Things (IoT) botnet framework, dubbed TuxBot v3 Evolution, has been discovered by cybersecurity researchers. The framework shows signs of being developed with assistance from a large language model (LLM), although the results are not entirely successful.

The LLM was used to generate botnet code, but it included a safety disclaimer that the developer failed to remove before shipping. This suggests that while the AI did aid in constructing the botnet, several functions in the analyzed samples failed to work correctly. A manual code review would have likely resolved these errors, and it’s possible that more polished iterations of the malware exist out there in the wild.

The TuxBot v3 Evolution framework consists of multiple components, including a C-based bot agent that cross-compiles for various architectures (ARM, MIPS, MIPSEL, MIPS64, x86_64, PowerPC, and RISC-V). The Go-based command-and-control (C2) server features a DDoS-for-hire panel, while the custom exploit virtual machine is designed to target vulnerabilities in IoT devices. Additionally, there’s an automated build system and Docker-based test infrastructure.

The bot agent is responsible for brute-force Telnet access on targeted devices using 1,496 credential pairs. It also incorporates exploit code targeting over 30 IoT device families using known vulnerabilities. The C2 server communicates with the bot agent over an encrypted TCP channel, employing a SHA512 domain generation algorithm (DGA) and peer-to-peer gossip protocol with Ed25519-signed commands.

The framework’s modular design allows for flexibility in its operations. It can resort to various fallback mechanisms, including Internet Relay Chat (IRC), DNS TXT queries, and HTTP polling. The lineage of TuxBot v3 Evolution has been traced back to three different botnets: Mirai, AISURU, and Wuhan. Some functions have also been ported from the open-source MHDDoS Python DDoS toolkit.

At least one sample of the malware was uploaded to VirusTotal on January 20, indicating it’s been around for over six months. Evidence suggests that work on the botnet began a year prior, when the author cloned the MHDDoS repository from GitHub. The framework’s description claims it features a professional-grade C2 platform with multi-user admin panel and automated deployment.

The Go-based C2 server component uses three different TCP ports for incoming connections: 1999 (or 31337), 2222, and 9999. These ports handle encrypted command dispatch to connected bots, interactive shell access over SSH, and programmatic interface via JSON, respectively. Once launched, the botnet follows a pre-defined initialization sequence.

This sequence includes loading the C2 address from a multi-tiered architecture with one primary channel and five alternate mechanisms. It also sets up anti-debugging and anti-VM protections to evade analysis tools. The process name is hidden, persistence is installed, and various sub-modules are launched to mount DDoS attacks and establish communication channels over IRC, HTTP, DNS, and P2P.

The dedicated HTTP scanner can manage up to 128 concurrent connections at any given point in time, operating with the goal of discovering vulnerable web interfaces. Persistence is accomplished through systemd service, cron entries, and a watchdog keepalive process to ensure TuxBot remains operational on compromised machines.

Multiple files contain raw LLM chain-of-thought reasoning left verbatim in comments. These comments reveal the internal reasoning as it worked through porting tasks, complete with self-interruptions, decisions, and references to ‘the user’ (meaning the developer who prompted the LLM). This suggests that the AI was used as a judge of sorts, providing guidance on how to develop the botnet.

The core working functions in TuxBot v3 Evolution, coupled with its reliance on AI tools for businesses, signal accelerated integration of features. The framework’s modular design enables what appears to be single developer to come up with a multi-pronged toolset featuring multiple C2 channels and custom exploit VM.

Shared infrastructure with Kaitori v3.9 and AISURU tooling places the TuxBot operator within the Keksec ecosystem, known for running multiple IoT botnet variants in parallel. This variant aims to go beyond Mirai forks by incorporating encrypted C2, DGA, and a modular exploit system.

The disclosure follows recent emergence of two other botnets: RustDuck and AryStinger. These have targeted routers, IP cameras, Android boxes, and poorly secured servers for co-opting them into networks designed to render online services offline and conduct reconnaissance.

OpenAI has developed a super-hacker model called GPT-Red, designed to test the security of its large language models (LLMs). The company claims that training GPT-5.6 against GPT-Red made it the most robust release yet. This is not just about making LLMs safer; it’s also about future-proofing OpenAI’s safety procedures and staying ahead of human attackers.

Artificial intelligence startup Anthropic is gearing up for a potential initial public offering (IPO) later this year, with bankers lining up investor meetings in advance of the share sale. According to an individual familiar with the plans, Anthropic’s IPO preparations are progressing rapidly as its advisors sound out demand from prospective investors before embarking on a formal roadshow.

The meetings suggest that Anthropic is moving closer to listing on public markets, which could happen as soon as October, although the exact timing remains uncertain. This would mark another significant milestone for the AI industry, following June’s massive SpaceX IPO and further opening up the public markets to companies at the forefront of the AI boom.

Anthropic confidentially filed its IPO prospectus with the Securities and Exchange Commission last month but has yet to disclose when it plans to debut on the stock market. The company was founded in 2021 by a group of executives and researchers who defected from OpenAI over concerns about the direction of their former employer, and has since found early success selling its AI assistants to enterprises.

Anthropic’s popular coding assistant, Claude Code, is one key factor behind its rapid growth, with the company closing a $65 billion funding round at a valuation of $965 billion in May. This puts it above OpenAI’s valuation for the first time and marks another significant milestone for the startup, which has been fueling the AI spending boom.

The involvement of top Wall Street banks Goldman Sachs, Morgan Stanley, and JPMorgan Chase in Anthropic’s IPO planning underscores their growing interest in the AI sector. The three firms are among the biggest on Wall Street by revenue, and their participation highlights the significant role that AI is playing in driving growth for financial institutions as they seek to satisfy investors clamoring for ways to fund or invest in this rapidly expanding theme.

Anthropic’s potential IPO would also mark a shift towards greater transparency and accountability within the industry. The company has been at the center of controversy over export controls on its technology, with the Trump administration recently lifting restrictions on certain products. However, it remains unclear how this will impact Anthropic’s future plans or operations.

The world of model routing isn’t as straightforward as it seems. Our experience in building routing into agentic systems has shown that what appears to be a simple problem quickly becomes a complex challenge.

One dimension that makes model routing surprisingly hard is cost. We expected GPT-4.1 to be cheaper than Claude Sonnet 4.6, but our tests revealed the opposite: Sonnet cost $79 total ($0.19/task) while GPT-4.1 cost $155 ($0.37/task), nearly double.

The explanation for this anomaly lies in caching – a factor that’s often overlooked when discussing routing. Agent workloads tend to reuse large chunks of context across steps, leading to high cache hit rates. When these rates are high, effective input costs drop dramatically.

This experience highlights the importance of considering actual cost, which depends on the interaction between the model, workload, and serving infrastructure. A router that only looks at pricing sheets is optimizing against the wrong numbers – a mistake with significant implications for data analysis tools.

Another dimension that complicates model routing is complexity. A common strategy is to estimate task difficulty and send harder tasks to stronger models, but this approach breaks down in two ways: difficulty can be invisible at routing time, and it’s just one signal among many that needs to be balanced with cost, latency, and other factors.

Secondly, even if task difficulty could be perfectly estimated, it’s not the only consideration. In production environments, routers need to balance multiple competing demands, including compliance requirements, data residency rules, privacy constraints, and approved model lists – all of which can influence where a task should go.

We’ve found that building effective routing systems requires considering more than just cost or quality: latency, reliability, and governance are also crucial factors. This complexity is often overlooked in discussions about routing systems, but it’s essential for data analysis tools that need to handle real-world applications with nuance.

Latency is another challenge when it comes to model routing. It’s tempting to think about latency solely in terms of model size – bigger models are slower, smaller ones are faster – but user experience depends on more than just model speed: infrastructure factors like hardware, cache warmth, and endpoint busy-ness often dominate end-to-end response times.

A theoretically faster model can still produce a slower experience if the serving conditions aren’t right. Routing granularity also plays a crucial role in determining latency: routing once per task adds minimal overhead, but routing at every step introduces additional decision points that increase latency and operational complexity.

These lessons have shaped our approach to building routers for agentic systems – we’ve shifted from treating routing as a classification problem to an optimization challenge. Our algorithm optimizes across cost, quality, and latency simultaneously while staying lightweight enough not to become a bottleneck itself.

The figure below illustrates the result on the AppWorld Test Challenge with a CodeAct agent. Each blue square represents a different configuration of our router, tracing out a cost-accuracy frontier – it shows that the router provides a range of operating points for choosing between cost, latency, or accuracy priorities.

Configuration 1 (latency-optimized) achieves 84% accuracy for $93 and 83s: a 21% cost reduction and 9% latency reduction compared to running Opus alone, with only a 4% accuracy drop. This is what our experience has led us to conclude about routing – it’s not really about choosing models; it’s about optimizing systems.

Models are just one variable – an important but not sole consideration alongside caching behavior, infrastructure state, compliance constraints, and workload patterns. When routing works well, it’s rarely because it found the ‘best’ model for a given task: it’s because it found the best operating point for the entire system.

That’s a harder problem than classification, but it’s the one worth solving – we believe that optimizing systems will lead to more effective data analysis tools. We’ll be sharing more about our technical approach in a follow-up post and look forward to hearing from others who are building routing into their own agentic systems.

Our experience has also shown us that standard difficulty-based routers often fall short, landing in similar accuracy ranges but at higher cost – they don’t explore the full tradeoff space that an optimization-based approach can. Our algorithm’s lightweight nature (roughly 6 ms and 2 kB of memory per task) means it doesn’t become the bottleneck we warned about earlier.

Our router optimizes across multiple factors, including cost, quality, latency, compliance, and reliability – all while staying adaptable to changing workloads and infrastructure conditions. This is a more complex problem than classification, but one that’s essential for building effective routing systems.

We’re sharing our findings in the hope of helping others who are struggling with model routing challenges – if you have any tradeoffs or complexities you’d like to discuss, we’d love to hear about them.

The complexity of model routing is often overlooked, but it’s a crucial consideration when designing data analysis tools. We believe that by prioritizing system optimization over model selection alone, we can build more effective and efficient routing systems – ones that truly meet the needs of real-world applications.

Protecting the world’s oceans requires accurate and reliable information. In high-stakes operational domains like maritime, incorrect answers can lead to significant resource waste and potential harm to personnel. The Skylight team at Ai2 aimed to build an AI agent that could provide real-time maritime domain awareness while ensuring reliability and trustworthiness. Shippy, their AI for this purpose, was designed with a focus on building a system that could be trusted to deliver correct answers within its limits.

Claude, the popular chatbot from Anthropic, has launched its new reflect feature to help users understand their behavior and optimize how they use the platform. This move comes as concerns about AI usage continue to grow.

The tool is designed to provide detailed insights into how users collaborate with Claude, highlighting key topics, usage patterns, and frequent tasks. Users can review chat activity over periods ranging from one month to a year to gain valuable information – like which features are used most frequently or where interactions tend to slow down.

One of the primary goals of the reflect feature is to encourage users to examine their relationship with Claude more closely. To achieve this, the tool periodically asks questions that prompt users to think critically about their AI use and engage in discussions with Claude.

The insights dashboard offers practical suggestions for using the chatbot effectively without providing repetitive information. The feature also incorporates Anthropic’s proprietary framework for evaluating AI use, which helps determine how to delegate tasks to a chatbot accurately – something experts say is essential when working with AI tools like this one.

Anthropic consulted independent experts in developing the reflect tool, focusing on creating insights that would help users evaluate what works best about Claude for them. The company also drew from youth developmental expertise to inform the design of tools that can assist young adults and parents in understanding their AI use – a key aspect as AI becomes increasingly integrated into daily life.

The reflect feature is currently available in beta form, with both free and paid subscribers who have enabled chat memories able to access it. Users must be at least 18 years old, but Anthropic has taken steps to ensure sensitive conversations are handled carefully – including measures against data breaches and unauthorized access.

Anthropic safeguards user privacy by not incorporating incognito chats into insight reports or accessing files from connected tools like email inboxes or health data. However, reflection insights may cover sensitive topics at a high level, and if Claude previously shared relevant resources during a chat, that information might appear on the dashboard – but only to provide context.

The introduction of this feature marks an effort by Anthropic to give users genuinely useful information about their AI use – something they hope will empower people to make informed decisions. This is particularly important when it comes to using AI tools in business and personal settings alike.

By leveraging data analysis, the company aims to help users interact with Claude more effectively. The goal goes beyond providing a snapshot of past behavior; it encourages users to think about how they can use these insights to improve future outcomes – whether that means streamlining workflows or better assessing AI outputs.

The reflect feature’s development is also an opportunity for Anthropic to collaborate with experts and inform the design of tools that can assist young adults, parents, and businesses in understanding their AI use. As a result, users will be able to access practical insights on how they interact with Claude – making it easier for them to make informed decisions about using these AI tools.

A rare opportunity has arisen for car enthusiasts to own one of the last Tesla Model S’s built, but it comes with a hefty price tag. A dealership in New Jersey is selling a 2026 Tesla Model S Plaid Signature Edition, one of only 350 examples produced by Tesla before ending production on its iconic model. The asking price? An eye-watering $259,995, more than $100,000 above the original sticker price of the limited-run car.

The Tesla Model S has been a staple in electric vehicles for over a decade, and its demise marks the end of an era. To commemorate this milestone, Tesla created the Signature Edition, which came with unique features such as Garnet Red paint, gold accents, carbon ceramic brakes, free lifetime Supercharging, and Full Self-Driving capabilities. These exclusive cars were only available to those who received an invitation from Tesla, and they had to sign a no-resale agreement.

The resale agreement is designed to prevent owners from flipping the car for profit by granting Tesla the Right of First Refusal to purchase it back at market value. If Tesla catches an owner selling the car without permission, it threatens to seek injunctive relief or impose a $50,000 penalty and put the original buyer on a do-not-sell list. This means that even if someone buys this rare car, they may lose out on two of its biggest perks: Full Self-Driving and Free Supercharging.

According to Tesla’s agreement, these exclusive features will not transfer with the vehicle when ownership is transferred. The new owner would have to purchase them separately or live without them. It remains unclear whether Tesla will actually enforce this language or impose the $50,000 penalty on the new owner. In the past, Tesla has blocked owners from transferring unlimited Supercharging access.

The question remains: is owning one of the last Tesla Model S’s worth an extra $100,000? On one hand, it’s a rare opportunity to own a piece of automotive history. On the other, previous limited-edition models have seen their value plummet on the used market. The 2012 Tesla Model S Signature Edition, for example, is now selling for well under $20,000.

Despite its hefty price tag, this car still offers impressive performance capabilities, with a 0-60 mph time of sub-two seconds. For those who want to own one of the last Tesla Model S’s and scratch their limited-edition itch, this might be their only chance.

A new study has shed light on the motivations behind people’s desire to automate tasks with robots. Researchers found that individuals are more inclined to automate activities based on their emotional experience rather than the time they consume. This challenges common assumptions about what drives automation preferences.

The investigation, which drew from three datasets - BEHAVIOR-1K, American Time-Use Survey, and its Well-Being Module - aimed to understand whether people prioritize automating tasks due to their duration or the feelings associated with them. The study’s findings suggest that happiness and pain are the strongest indicators of automation preferences.

Interestingly, time spent on activities does not strongly predict automation choices. This means that individuals may be willing to invest significant amounts of time in certain tasks if they derive pleasure from them, but opt for automation when those same tasks become painful or stressful.

The study also identified differences in automation preferences across various social groups. Women tend to prefer automating stressful activities, while men prioritize automating tasks that make them unhappy. Mid-income individuals, on the other hand, are more likely to automate less enjoyable and meaningful activities.

Low- and high-income individuals showed no significant correlations between their income levels and automation preferences. The researchers hope that this study will inform the design of robots that align with user priorities, leading to more socially relevant solutions in domestic robotics.

A New Jersey dealership is trying to sell a Tesla Model S Signature for an eye-watering $259,995. This price tag represents a staggering markup of over $100,000 compared to what Tesla originally charged buyers for the collector’s edition.

The car in question has just 297 miles on its odometer and carries VIN #71, indicating it is one of only 250 units produced as part of an invite-only farewell run for the Model S and Model X. The listing at J&S Autohaus in Ewing, New Jersey, also includes a $495 documentation fee, bringing the total price to $260,490.

The Signature Series was built as a farewell run for the Model S and Model X, with only 250 units produced. Each car features exclusive Garnet Red paint, gold brake calipers on carbon-ceramic brakes, and 21-inch wheels. The low VIN ending in ‘S00071’ also confirms its status as one of these highly sought-after collector’s items.

Tesla originally priced the Model X Signature at $159,420 and the Model S Signature around $155,000. This means that the J&S Autohaus listing represents a premium of over $100,000 compared to Tesla’s own price for the car - before it has even been delivered to the buyer.

When you compare this markup to what Tesla charged for the standard Plaid model, which was around $124,900, it looks even steeper. The reason behind this high asking price is likely due to scarcity and exclusivity.

The Signature Series was invite-only, with only a select group of existing owners receiving an email from Tesla offering them the opportunity to purchase one of these limited-edition vehicles. This combination of factors has attracted flippers hoping to cash in on collectors who missed out on this chance.

The ‘Signature’ name also carries historical significance. When the Model S first launched in 2012, the initial ~2,000 cars sold were Signature editions that required a $40,000 deposit and cost nearly $100,000 each. These early owners were Tesla’s original believers.

However, beneath its exclusive paint job and gold trim, the Model S Signature remains fundamentally the same car as the standard Plaid model. It lacks new battery cells, faster charging capabilities, steer-by-wire technology, or any meaningful range gain.

The buyers are essentially paying a six-figure premium for cosmetic upgrades and a numbered plate. The Tesla community is divided on this issue, with some enthusiasts defending the high price tag due to its exclusivity and historical significance.

Others argue that it’s nothing more than a marketing gimmick designed to separate collectors from their hard-earned cash. Even the author of this article owns one of these early Signatures and doubts it would fetch anywhere near $260,000 today.

It remains to be seen whether anyone will actually pay six figures for one of these limited-edition vehicles - the original 2012 Signature editions were genuinely historic but did not appreciate into collector cars. In fact, they’re unlikely to sell for anything close to this price tag in a private sale.

If you’re considering investing in a Tesla Model S, there may be better options available. With electricity rates climbing nearly 10% last year, home solar can protect against future rate increases - and with lease and PPA options, you can go solar with zero upfront cost and start saving immediately.

A new state-run artificial intelligence (AI) platform has been launched in California to support government workers. The platform, called Poppy, is a collection of 10 different models tailored to fit the diverse and challenging work of government. This AI assistant was made possible by Executive Order N-12-23 signed by Governor Gavin Newsom late last year.

The development of Poppy involved overcoming several challenges, including privacy, security, governance, and appropriate use considerations. Despite these hurdles, the California Department of Technology (CDT) successfully implemented a consolidated suite of models in just under a year. This achievement is significant given that getting any model up and running within state government can be a daunting task.

According to Shera Mui, deputy director of CDT’s Platform Services, what started as an effort to create a statewide chatbot quickly evolved into something more useful – a platform that could be managed and adapted to new technologies in real-time. ‘We discovered early on that we didn’t want to create large language models; that was not going to be the smart thing to do,’ Mui explained.

The team at CDT took a different approach, focusing on creating a platform that would allow users to securely choose a model best suited for their task. This flexibility is particularly useful in government work, where tasks can range from deep analysis of legal documents to creating code for new websites. The platform’s ability to keep all user data and information secure has been a major driving factor behind its development.

The early access phase involved giving 70 departments with around 100 users each the opportunity to test the first iteration of Poppy. This pilot program helped drive discussion around potential use cases, which in turn informed the development of subsequent models. The platform’s flexibility is one of its key strengths – it allows users to choose from multiple models without having to duplicate data or compromise on security.

Dr. Lucy Andrews, a scientist with the California Department of Water Resources (DWR), has seen firsthand the impact that Poppy can have in labor- and time-intensive work. She noted that the tools have played a critical part in determining the intellectual and financial value of her department’s contributions across various sources. This is particularly important for DWR, which invests tens of millions of dollars annually in scientific research.

Previously, tracking these contributions was a manual process that involved weeks of data entry using an Excel spreadsheet. With access to AI tools like Poppy, this time-consuming task has been significantly reduced – users can now complete it within 1-2 hours with a slow internet connection. The ability to track and analyze scientific relationships more effectively is also a major benefit for DWR.

Andrews noted that the department’s focus on accountability means they are taking a conscientious approach to model selection and use cases before implementing Poppy. This includes considering the environmental impact of running data centers, which support AI operations. While there may be some concerns about energy consumption, staff is working to minimize this by carefully selecting models and optimizing their usage.

Early metrics from the statewide launch show that users are interested in document summaries, policy analysis, and complex code creation – all key areas where Poppy’s capabilities can make a significant difference for government workers. Some departments have also expressed interest in exploring image generation features as part of future additions to the platform. However, more work needs to be done on outlining a long-term roadmap for these developments.

As for what’s next for Poppy, Mui sees opportunities for further expansion and improvement. This includes adding new features, tailoring models to specific departmental needs, and streamlining data sets to enhance cooperation among departments. CDT has requested $1 million in funding during the budget process toward this end – a significant investment that reflects the potential of AI tools like Poppy.

It’s worth noting that while Poppy is an advanced toolset for government workers, no decision is left solely to an AI system. Humans maintain control over all outcomes and are responsible for ensuring that AI-generated information aligns with their goals and objectives.

A previously undocumented modular internet-of-things (IoT) botnet framework named TuxBot v3 Evolution has been identified by researchers. This malware leverages a large language model (LLM) to assist in its code development, resulting in mixed outcomes. While the AI generated botnet code as requested, it included a safety disclaimer that was not removed before deployment.

Sellers and managers are finding ways to streamline their workflow with the help of AI-powered tools. One such tool is ChatGPT Work, which enables sales teams to turn scattered account information into actionable plans. This involves consolidating data from various sources like CRM fields, call notes, email threads, and customer documents.

Sales work often gets fragmented across multiple platforms, making it challenging for teams to keep track of everything. However, with ChatGPT Work, this context can be pulled together quickly, producing a usable version of the artifact in no time. This could be an account brief, meeting prep packet, forecast review, or even a stalled-deal diagnosis.

The tool helps sales teams bring customer context into their work by identifying high-priority accounts and signals. It also enables them to prepare for meetings, complete follow-ups, update customer records, build close plans, and review deals at risk. ChatGPT Work integrates with popular tools like Salesforce, HubSpot, Slack, Outreach, Clay, Rox, and Actively.

Already using ChatGPT Work? The sales plugin is available for installation on the platform. For those new to the tool, there are use cases available that demonstrate its capabilities in real-world scenarios. By leveraging AI-powered assistance, sales teams can focus on high-value tasks while leaving routine work to the machines.