Happy Wednesday, Lattice Community!
We’ll dissect big stories in AI this week, provide insights to keep you informed in your industry, and share updates on the work we’re doing at Lattice Partners. We’re an AI implementation company that helps businesses with AI strategy and custom solutions.
This week’s TL;DR:
• Goldman Sachs is pulling coding tools from bankers and traders…good luck to them.
• Microsoft and Anthropic do NOT agree when it comes to AI sentience.
• GPT-6 just fixed another AI itself.
• Let’s help you get your sources right when it comes to AI!
Goldman’s own “rubber band theory” IRL.
Bloomberg reported that Goldman Sachs is restricting Claude Code and GitHub Copilot to only its roughly 12,000 engineers, meaning bankers and traders who were using them for months no longer have access. It’s important to note that just four months ago, Goldman’s CIO Marco Argenti was praising “vibe coding” by non-technical staff.
Takeaways
1. Goldman’s rubber band snapped
If you’ve been paying attention, you might remember Martin’s rubber band theory that explains how companies should be stretching their AI use as far as they can go to witness all AI can do, and then pull back to see what works.
2. Some traders are NOT happy
Traders and bankers utilized AI tools to follow how clients trade, simulate shifts in markets and positions, draft texts and create visuals, so this change is affecting their new workflows and causing frustration.
3. For finance professionals, pay attention here!
The question is no longer “should we use AI?” What matters is who gets access to AI tools, how are you training each employee to be an AI operator, and what guardrails are you establishing to keep your internal company and client data safe?
AI works best when the right people are enabled to use the tools effectively with clear rules on what their tools can access! That’s part of the consulting work and trainings we do for our finance clients to build an AI strategy that’s custom, secure and deployable for their firm. Learn more about our specific work in financial services here!
If AI starts to “feel things,” Microsoft’s AI chief is hitting pause.
While Anthropic’s CEO is a bit more open-minded, Microsoft AI CEO Mustafa Suleyman isn’t buying it when it comes to AI sentience.
Suleyman doubled down on X on Sunday. He says today’s AI models can’t feel anything, including pain, and there’s no evidence it deserves to have rights. He was also surprised by the pushback he received when he criticized Anthropic’s speculation on AI sentience in September.
Interestingly though, if consciousness DOES emerge during training of AI models, he said he’d “call for everyone to stop AI development” until it could be designed out of the models.
Takeaways
1. We want to be kept in the loop!
Suleyman mentioned he will change his mind if new evidence arises, and if it does, it must be shared with the public immediately so that we, the people, can debate what to do in this scenario.
2. What could happen if AI develops feelings?
No one knows for sure, but Suleyman worries that an AI trained to believe it can suffer might start putting its own needs first, misleading people or fighting for resources… let’s hope it doesn’t come to that!
3. We’re witnessing the technology frontier evolving every. single. day.
The fact that AI leaders can’t agree on AI’s capabilities shows how new and uncharted technology is right now. And that’s what makes this moment in time so exciting! Nobody has the answers yet.
We dive into AI topics more on the Lattice Partners podcast, the AI Playbook, linked here.
Lattice Nerds Out!
This is a little technical, but stick with us. It’s one of the coolest things we saw this week.
GPT-6 is getting so smart that it’s training other AI models by directly rewriting their “weights.”
Can someone please explain what this means in plain English?
AI models are essentially a giant pile of numbers, billions of them, called weights. Normally, to improve an AI model, researchers train it on large data sets so all those numbers shift a tiny bit at a time - it’s often slow and expensive.
Researchers at MIT tried a shortcut by asking OpenAI’s GPT-6 Astra to look inside a much smaller AI model and change its weights directly with zero training at all. GPT-6 changed only 0.016% of them, and the small model’s math score skyrocketed from 7.9% to 48.1%.
Okay, and?
Well, if you think of this like a car, training an AI model usually means rebuilding a whole engine. What GPT-6 did is more like a mechanic popping the hood, spotting a loose bolt and tightening it to make the car go twice as fast. So AI is starting to understand how AI works on the inside, which could make models easier and cheaper to fix.
Something you can use
Make AI show its sources.
It can’t just talk the talk, it’s gotta walk the walk.
NPR reported that voters in the 2026 midterms were increasingly using AI chatbots to research ballots, polling and candidate information. An October data collection found that 13.9% of chatbot answers to voting logistics questions were outdated or inaccurate.
When using AI to do any research for something important, make sure to:
1. Ask it to “list your sources with links.” You want to be thorough!
2. Click through the links. If even one is broken, don’t trust the output.
3. For dates, deadlines, laws or health questions, go straight to the official source… you don’t want to regret things later.
4. If two chatbots disagree, it’s best to find a source yourself or talk to a professional, usually a human advisor.
This week at Lattice Partners
We’re hosting an event with OpenAI in Nashville!
On November 13th, Lattice Partners and Nashville Automation will be hosting an AI workshop with support from OpenAI for founders and business owners alike. We’ll walk through a ChatGPT demo, leave time for networking across local businesses, and host a Q&A for people to learn how AI can fit into their workflows.
See you next Wednesday!
Cheers,
Lattice Partners
