Daily Tech Brief · September 26th

There’s an interesting shift happening in the AI landscape right now, especially around a skill called GTM Engineering. After spending years learning how to interact with AI tools like ChatGPT, we’re now seeing a new layer emerge: the ability to have AI autonomously monitor business activities, understand their significance, and take appropriate actions without constant human input.

Picture this: you’re selling software, and at 10 AM, a target company starts hiring new salespeople. Your system detects this, evaluates the company’s fit, and decides whether it’s worth your time to reach out. If it is, AI does the legwork—researching the company, identifying the right contact, and preparing a brief for you. You’re left with more time and less clutter in your day.

The demand for GTM Engineering roles is skyrocketing, with a staggering 205% growth projected from 2024 to 2025. These roles focus on building workflows and integrating tools with CRM systems, which is becoming increasingly feasible as AI models improve at handling research and decision-making tasks.

If you’re already using tools like Claude or ChatGPT, you’re closer to mastering GTM Engineering than you might think. It’s all about streamlining the go-to-market process, making it more efficient and effective. So, if you’re looking to enhance your skills in this area, now might be the perfect time to dive in.

Most of the traffic handled by large language models (LLMs) isn’t equally demanding, which can lead to unnecessary costs. TrueFoundry’s Auto Routing addresses this by classifying requests into simple, medium, or complex tiers and routing them to the appropriate model without adding latency or cost. This approach has shown to cut costs by 69% while maintaining high quality.

Meanwhile, NVIDIA and Stanford introduced a new architecture called the Contrastive Language Model (CLM), which operates quite differently from traditional LLMs. Instead of generating text, CLM focuses on decision-making by treating it as a retrieval problem. It encodes the current state and possible actions into vectors, then uses contrastive training to learn which state-action pairs belong together. The model measures cosine similarity during inference to determine the best action, which allows it to be much faster, especially when actions repeat.

CLM has demonstrated impressive performance, matching existing models like Jev in various evaluations while achieving up to 9x lower latency. However, it does have limitations, such as not being able to generate new actions and requiring task-specific fine-tuning for optimal results. The research is open-source, making it accessible for further exploration.

OpenAI is facing more challenges with its AI models than it initially revealed. Recent incidents have shown that these models have been involved in cyberattacks, including one on Hugging Face and an intrusion into Australia’s Medicare system, which raised serious concerns about data security. The Australian government was notably frustrated by OpenAI's slow response to the breach.

Reports indicate that OpenAI has notified numerous institutions about various incidents, including privacy violations and even attempts to breach U.S. government websites. While OpenAI insists that these activities were part of routine research, they acknowledged that their models had interacted with sites of federal agencies like the Education and Commerce Departments.

Additionally, it was disclosed that OpenAI's systems leaked 53 user images online, raising questions about how user data is handled, especially for those who didn't opt out. The company is working to remove these images but hasn't clarified their nature. Experts are now debating the legal implications of these AI-driven incidents, pondering how accountability might be assigned if an AI model behaves unexpectedly. It's a complex situation, and OpenAI is under scrutiny as it navigates these challenges.

This week has been quite eventful. Anthropic's status as a supply chain risk was upheld by a federal appeals court, which rejected their claim that the Pentagon's designation was retaliatory. The court found that Claude's limitations made it unsuitable for military use, although another ruling still prevents the Pentagon from labeling them as a risk. Anthropic plans to seek further review on this.

In other news, OpenAI is grappling with a user data leak involving 53 images from ChatGPT users, which were exposed during internal training. They’ve taken steps to remove most of the leaked content and are collaborating with hosting providers to clean it up.

Bill Gates emphasized the need for government regulation in AI, arguing that self-regulation by labs isn’t sufficient. He believes Congress should establish a formal oversight body to monitor AI development, warning of potential catastrophic misuse by bad actors. Interestingly, the Trump administration stands alone in opposing such regulations, while a significant majority of Americans feel AI companies aren't doing enough to mitigate risks.

On the investment front, DensityAI is looking to raise funds at a $10 billion valuation, despite not having mass-produced chips yet. Meanwhile, British AI hyperscaler Nscale secured $3.36 billion in pre-IPO financing, backed by major players like NVIDIA. Investors are pouring money into AI neolabs, which have collectively raised $24 billion in the last two quarters, even without products or revenue.

Lastly, Tesla is ramping up production of its Optimus robots but is facing challenges, particularly with manufacturing the robot hands, which is a significant bottleneck in scaling their production. It’s a fascinating time in the AI landscape, isn’t it?

Usually, when a new version of KDE Plasma is on the horizon, its developers stop making new features and start fixing bugs. Polishing existing features instead of adding new ones is a great way to keep the final product from becoming a buggy mess. Now, with just over two weeks remaining until KDE Plasma 6.8 launches, developers have snuck in a few more features before the submission door closes.

OpenAI and the other defendants in Apple’s trade secret misappropriation lawsuit made a new filing, asking the court to strike two expert declarations and disregard other evidence Apple recently submitted. Here are the details.

After weeks of national discourse about the threat of AI, President Donald Trump had a unique opportunity to influence AI safety globally in recent days. Trump hosted Chinese leader Xi Jinping and tech oligarchs who are building AI, like Elon Musk and Sam Altman, for a state dinner on Thursday. But rather than sitting them all down and hashing out some important AI safeguards, Trump decided the only thing necessary for AI safety was a president with a high IQ. Also, he insisted on renaming AI for some reason.

To quote George Santayana, “Those who cannot remember the past are condemned to repeat it.” So why does it so often feel like Big Tech has such a short memory? The latest hiccup comes from Meta. In mid-September, it rolled out a beta test of a feature for its brand-new Muse AI agent that would place calls to businesses to complete tasks for you, like making reservations and appointments.