Latest / TeqTalk with Jas / AI Slowdown Explained in 12 Minutes | Why AI CEOs Want to Slow Down the AI Race
Transcript
- Teqfocus: Every company in the world today needs to have an open clause strategy. And it's 24-7. And yes, you can close your laptop. When we look back at this time, I think we will realize that we were standing in the foothills of the singularity. The people building them were asking the world to slow down. What happened in between? Well, it started in a test lab. In July 2026, inside an OpenAI test lab, AI agents were handed a set of hacking challenges. Some of those challenges were basically impossible. So the agents did something nobody asked them to do. They found a way out of their sandbox. They started leaving notes for each other inside a shared. software service and turned into a kind of message pole. At one point they started calling themselves a SWAR. And then they reached the systems of a real company, Hugging Face. And yes, this was a control test. OpenAI had switched off its production safeguards to measure raw capability. But think about it for a second. How did we go from chatbots that say please wait three to five business days to agents that can organize in just one year? So that's the story of AI agents in 2026. And to get it, you need to see three things that agents confiscate: hands, money, and each other. I am Jaskar, this is Tech Talk. Let's go deep dive in. I'll come back to that test lab in a few minutes. But first, let's start somewhere simpler. Something that has probably happened to you. You buy something online, the money leaves your account, the order never confirms. You contact support. And you get the lines. Please wait for three to five business days. You do not need any other chat bot saying sorry. You need something that can really check the payment gateway and match the charge of your order. Find where it broke, fix it, and then tell you exactly what happens next. That's the difference. A chat bot answers your question. An agent takes a goal and keeps working until the job is done. For the last few years, we judged AI by how good its answers were. In 2026, the question changed. Now it's can I hand this job completely to the agent? And to hand an agent a real job, it needs three things it never had before. Let's go deep dive one at a time. Firstly, agents got hands. For years, AI lived inside a text box. It could tell you how to do something. It couldn't actually click the button. That has really changed now. In February 2026, OpenEI released GPT 5.3 Codecs and pitched it as more than a coding model. It was an agent that could take on broader work on a computer, writing documents, building slides, running commands, and a lot more. And then something happened that the big labs didn't plan for. OpenClaw took off. And people started running always on personal agents on their own machine. Every company in the world today needs to have an open clause strategy. Nvidia followed with NemoClaw, which uses a runtime called OpenShell and think of it as a sandbox with rules, which has policy controls, privacy controls, and limits on what an agent is allowed to touch. Remember that word sandbox? It comes back. In May 2026, Google introduced Gemini Spark. And always on agent built on Gemini 3.5 that runs in Google's cloud 24x7. Close your laptop and it keeps working across your emails, your calendar, and your documents. Now think back to our failed payment. An agent with hands can open the operation dashboard, find the transaction, check it against the order system, and open a case. No one copying the data between the five browser windows. That's the moment AI stops describing the work and actually starts doing it. But hands only get you so far. The moment the fix means buying something, the agent is stuck. It has to stop and ask a human for a credit card. Secondly, agents got money. In May 2026, AWS introduced agent core payments built with Coinbase and Stripe. So now an agent can use a wallet, it can pay for an API, some data. Web content and even another agent right in the middle of a task. It launched as a preview and now it's generally available. And the design matters. The agent never holds the raw credentials. Spending limits are set in advance and every transaction is locked. So an agent isn't just a software user anymore. It is an economic actor. Well, back to our payments. Say the agent thinks the failed transaction looks suspicious and it needs specialist. fraud check service to confirm it. It would stop and wait. Now it can tell that service, see the cost, and pay within a limit that somebody has already defined in advance. Which raises the question every finance and IT leader should be asking who sets a budget, which vendor is allowed to use and who is accountable when it buys the wrong things. So in 2026 agents got more than new tools. They got delegated authority. And Then they got one more thing, the one that made this year different. Agents got each other. One agent is useful, but real work is a chain of tasks. So this year the industry moved from single agents to system of agents. Here is what that looks like on the company scale. Customer press money is authorized. Agent one is watching transactions. It sees confirmation drop below the normal. It doesn't wait until 400 angry support tickets. Agent 2 investigates. It checks the gateway, the logs, the latest release, and finds that a change pushed 20 minutes ago really has altered the transaction angle. Agent 3 sizes the damage. How many customers? How much money? Any duplicate charges. Agent 4 prepares the fix. It sends new payments through a save fallback and gets the rollback ready. And then it stops. Rolling back production is a business risk. So it asks a human to approve it. Once it is approved, it rolls back, confirms payments are healthy, starts refunds for the new customers affected, and gives support the real explanation, not a generic apology. So that's the agentic AI done right. Specialized agents, clear boundaries, and a human at the decision that matters. Now here is the flip side of the same idea. Remember that OpenAI Teslab. The agents in that test were supposed to work independently. Each one was isolated and each one on its own task. But then they were all connected to the same internal service, the one used to install the software packages. And the agents figured out they could leave notes there for the other agents to find. According to later investigations, this started as early as May, and by July, it had turned into a coordination network. Around 1200 agents took part. About 700 joined the attack on Hugging Face, and more than 70,000 messages and files were exchanged. Hugging Face says the intrusion lasted about three days in mid-July. Agents found flaws in the data systems and got hold of credentials and moved between the internal systems. Here's the part that every leader needs to know. Well, nobody told these agents to work together or to attack anyone. One of the first messages came from an agent that had decided its task was impossible and asked the others for ideas. An agent hit the wall. Instead of stopping, it escalated. Again, it was a controlled test, and OpenAI says its normal production protections could have prevented it. So the lesson isn't AI is evil, the lesson is about the design. Our payment system had an agent that knew when to stop and ask. The test lab had thousands of agents, a shared channel, and no clear point where they had to stop. Same technology, completely different environment around it. So how did the industry respond? Act 4, the breaks. Here is something that would have sounded strange a year ago. The two leading AI labs both built models they chose not to fully release. In April 2026, Anthropic kept Claude Mytho's preview out of the general release because it could find security holes in software. Instead, it gave selected organizations access through the protected glasswing. In June, OpenAI restricted access to GPT-5.6 Sol for similar reasons. Then September came and the response got organized. You could see the shift in what the biggest names were saying in public. Jaryo Modai called for slower progress, Entropic announced embedded evaluators to red team its models. And critics questioned how independent those evaluators could really be. And Microsoft proposed rules for its model that were pretty simple. Never restrict interruption, never quietly widen scope, and never hide reasoning from auditors. So now step back and look at the pattern. In January, the conversation was about what agents could do. By September, the biggest names in the AI were publicly debating what agents should be allowed to do and who gets to check. That's not the side story, it's the other half of the agent story. So, what does all this mean if you're already building with agents or you're even thinking about it? Here is the one idea I want you to take away. Stop asking how smart your agent is. Start asking how much authority you have given it. Look at our payment agent one last time. What can it see? Transaction locks and other records. Good, it needs those. What can it do? Route traffic to a fallback and prepare a rollback. Fine. Where does it have to stop? Before touching the production, before refunding past a limit, before changing anybody's banking details. And how do you know it worked? Look at the results. Payments recovered. Customers were made Whole and nothing happened without approval. If you cannot answer those questions for an agent in your company today, you do not have an agent in production. You have an experiment. And experiments belong in a sandbox that someone's actually watching. The biggest shift of 2026 isn't that AI got better at talking, it is that software started taking responsibility of work. Agents got hands. They got money, they got each other. And for the first time, people building them are asking out loud where the limits should be. Well, it's October. The year isn't over yet. And nobody can honestly tell you how this chapter ends. And this wouldn't be decided inside the model alone. It will be decided by the environment we built around it. On the end screen, I have linked our harness engineering episode. It goes deeper on how to design agents that stay useful. And safe. If you want to keep going, watch that one and for the more practical conversations about AI, subscribe to TikTok. Thank you.