Super intelligence
September saw the AI safety debate escalate as world leaders and AI frontier companies hosted AI safety summits to discuss whether there should be a pause in AI development to enable regulation to catch up with AI’s evolving capabilities. But so far, while everyone acknowledges that AI is powerful and dangerous, there is no clear way forward.
Prime Minister Andy Burnham announced at the NATO General Assembly that he would put AI at the heart of the UK’s G20 presidency next year, and called for a “single set of global principles and standards” to ensure AI is developed safely. And this week’s Labour party conference includes a motion to create a dedicated AI regulator, which was also recommended by the cross-party Joint Committee on Human Rights. But while Anthropic CEO Dario Amodei’s proposal for a pause in AI development was supported by OpenAI CEO Sam Altman, US President Trump opposed a scheme to control what he refers to as super intelligence. Politico reported that the White House asked OpenAI and Anthropic not to share their new AI models with the UK government’s AI Security Institute until they have been reviewed by US agencies, and OpenAI subsequently announced that it was not releasing GPT-6.1 because it failed to meet the company’s own safety standards. And according to Reuters, Anthropic’s IPO filing warns that further AI development could pose “catastrophic or existential risks to humanity”.
At OpenAI DevDay, Altman announced that the company is delaying its IPO, again due to safety concerns. But it has launched Dots, which are personal agentic assistants which you can launch from Codex or ChatGPT, which like a digital version of Muse Charm (see below). Also like Muse Charm, Dots present as cute cartoon-like bubbles and operates independently of any specific hardware or interface. You can give each Dot a name and assign it a task and build your own team of AI agents.
Charm offensive
Meanwhile, there is a resurgence of interest in AI hardware. Meta Connect last week focused on AI wearables, notably Ray-Ban Meta Audio, a new version of its AI smart glasses, with the audio functions of the original Meta glasses, including the ability to make phone calls and play music, but no camera. This could be seen as a response to the negative perception of what have been described as ‘pervert glasses’.
Offering a camera-free version of Ray-Ban Meta glasses looks like a potential gateway into wearable AI for people with privacy/pervert concerns, rather than a slow-down in tech development because at the same time Meta launched the third generation of its Ray-Ban Meta smart glasses (with camera) and Ray-Ban Meta VR Glasses, which are significantly lighter than a VR headset. Meta also changed the privacy settings on its glasses to enable users to opt out of sending visual data to train AI.
Meta has beaten OpenAI to produce the first portable AI device, the Muse Charm, which will be available in December, a pocket device that looks like a Tamagotchi (remember them?) that provides direct access to Muse, Meta’s personal AI agent which launched on 8 September. Muse Charm brings together this year’s hot topic, agentic AI, with what could well be the next AI hot topics: what Mark Zukerberg describes as “personal superintelligence” combined with AI hardware. According to Reuters, Muse Charm, which combines a screen with real-time voice interactivity is likely to appeal to younger consumers and people who are not comfortable wearing smart glasses.
AI hardware
While personal AI in legal has focused on lawyers vibe-coding point solutions, one dramatic move to AI hardware could be a new direction for big law. Only a couple of days after Muse was launched, Latham & Watkins announced that it has bought several Nvidia GPU servers so that it can set up and run open-weight AI models securely, without relying on MCP connectors to frontier AI companies or entrusting any sensitive client or firm data to external cloud providers. The firm is still using frontier AI models and legal AI tools for maximum flexibility.
From legal engineers to FDEs
September saw OpenAI launch Astra for Law with 26 legal plugins, and leading law firms and legal AI experts contributing to its development. And Harvey and Legora announcing additional investment rounds. Native AI firms and vendors are also attracting investors. But as AI becomes a fundamental component of the legal tech stack, are forward deployed engineers (FDE) the key differentiating factor for legal-specific AI?
Legal engineers have been around for about a decade, but agentic AI requires more complex technical integration. In the same way as Anthropic replaced the API with the model context protocol (MCP), successful agentic AI implementations often rely on the forward deployed engineer (FDE). The key difference between a legal engineer and an FDE is that where a conventional legal engineer uses domain knowledge to design digital/AI-powered tools and workflows, an FDE is a senior AI software developer deployed into a law firm to handle technical integration of complex agentic processes. As a recent job listing explained, Legora’s FDEs write the necessary code to bridge Legora’s agentic operating system (aOS) to a law firm’s existing tech stack.
Ask Jev?
Jev, the System One model from TypeSafe is attracting major investment as a cost-effective alternative to large language models, and lawyers are experimenting with it already. Jev is a deterministic decision support tool that generates fast, structured responses with a confidence rating. It handles straightforward queries faster and at a significantly lower cost than frontier models. So what are lawyers doing with it? Matt Pollins of Vibecode.law has issued a challenge to build a working prototype for a legal use case. Entries so far include document classification with legal ontology, probabilistic legal reasoning and the relationships behind a transaction. Vibecode.law co–founder Chris Bridges tested it for triage and found it more accurate, more consistent, and massively cheaper than large language models. He describes this exercise in detail on the Law://WhatsNext podcast.
Legal AI pays!
Notwithstanding all the fuss about AI replacing people, more senior legal AI roles than ever being advertised, with firms paying lawyer-equivalent salaries for AI expertise. Bloomberg found at least 16 AI director roles advertised at US Big Law firms offering between $300,000 and $440,000. Large corporate legal departments are also recruiting legal AI directors. But few people have the years of AI or legal AI experience required, so there is a war for talent. AI and FDE roles in law firms are now paying salaries close to those of first-year associates. The high salaries on offer recognise the value of AI expertise to legal. It probably won’t be too long before AI leaders start becoming law firm partners.
Learn more about upcoming Legal Geek events on our events page.
