The multimodel switch

Welcome back to The Shift. Let’s get straight to what matters in AI today…

Today we have:

🐡 Sakana AI Bets On Many Models Instead Of One

⚡️How to Put Recurring Tasks on Autopilot With Viktor

🛡️ DeepMind Is Preparing For AI Agents Gone Wrong

🔨Tools and Shifts you cannot miss

🐡 Sakana AI Bets On Many Models Instead Of One

As export controls create uncertainty around access to frontier AI systems, Japan's Sakana AI is taking a different approach. Its new Fugu platform combines multiple models behind a single API, aiming to deliver top-tier performance without relying on one provider.

The Shift:

One Request, Many Models - Fugu automatically selects specialist models, assigns subtasks, verifies outputs, and combines responses, allowing users to access a coordinated multi-agent system through a single endpoint.

Two Performance Tiers - The platform offers a lighter Fugu model for coding and everyday tasks, alongside Fugu Ultra, which targets more demanding workloads such as patent analysis, research, and security testing.

Chasing Frontier Performance - Sakana claims Fugu performs at or above Fable 5 and close to Mythos Preview across several coding, reasoning, and scientific benchmarks despite relying on model orchestration.

Built For Reliability - The company argues that distributing work across multiple models reduces dependence on any single frontier provider, helping shield users from disruptions caused by export controls or model restrictions.

Fugu reflects a growing belief that the future of AI may not belong to one dominant model. Instead of building a larger model, Sakana is betting that intelligently coordinating multiple systems can deliver frontier-level capabilities while improving resilience, flexibility, and access.

Together with HubSpot

LLM traffic converts 3× better than Google search

58% of buyers now start their research in ChatGPT or Gemini, not Google. Most startups aren't showing up there yet.

The ones that are get cited by the AI tools their buyers, investors, and future hires already use. And they convert at 3×.

Download the free AEO Playbook for Startups from HubSpot and get the exact steps to start showing up. Five minutes to read.

Get the free playbook

⚡️How to Put Recurring Tasks on Autopilot With Viktor

The work that eats your week shouldn't need you to start it every time. Here's how to set it up once and never think about it again.

Step 1: Add Viktor to Slack or Teams - Install Viktor from the app directory. No IT ticket, no setup project. It shows up like a new teammate in under two minutes.

Step 2: Connect your tools - Link your CRM, analytics, Drive, calendar, or whatever the recurring task touches. Viktor connects to 3,200+ tools and acts on them, not just reads them.

Step 3: Assign the task - Tell Viktor what you need done in plain language. Pull weekly numbers from Stripe and HubSpot. Send the Monday recap to the channel. Flag anything that moved more than 10% week over week.

Step 4: Ask it to run on a schedule - At the end of the task, tell Viktor when to repeat it. Every Monday at 8am. Every Friday before close. It confirms and runs it automatically from that point forward.

Step 5: Come back to finished work - The output lands in your channel on schedule. Report, dashboard, digest, whatever you asked for, done before you ask.

Viktor asks for approval before anything irreversible. Everything else runs without you touching it.

Free to start. Up to $100 in credits, no credit card required. You can get started for free here.

🛡️ DeepMind Is Preparing For AI Agents Gone Wrong

Google DeepMind has released a new AI Control Roadmap focused on securing advanced AI agents. The framework assumes that even well-trained systems can fail, making operational safeguards just as important as model alignment.

The Shift:

Treating Agents Like Insiders - DeepMind compares advanced AI agents to employees with privileged access, recognizing that systems capable of reading files, writing code, and using tools could create serious risks if compromised.

AI Watching AI - The roadmap proposes trusted AI supervisors that monitor an agent's reasoning, plans, and actions, creating an additional layer of oversight beyond simply training models to behave correctly.

Controls Match Risk - Safety measures would scale with the potential impact of an action, ranging from delayed reviews for low-risk tasks to real-time intervention and blocking for high-risk behavior.

Safety Gets Measured - DeepMind outlined concrete metrics, including coverage, which measures monitored activity, recall, which tracks detected bad behavior, and response time, which measures how quickly issues are stopped.

The challenge is no longer chatbot safety. AI systems are increasingly taking actions, not just generating answers. DeepMind's roadmap treats AI security as an operational problem, borrowing concepts from cybersecurity to ensure powerful agents can be monitored, audited, and controlled before mistakes become real-world consequences.

Together with Viktor

Your creative brief is due Friday. Viktor wrote it Tuesday.

Tell him the campaign. Viktor pulls last quarter's performance from Meta and TikTok, scrapes competitor ads, drafts the brief, posts it for review. You edit, he ships the creative requests to your designer. Inside Slack.

Get Started for Free

🔨AI Tools for the Shift

📊 RowSpeak – Lets you upload spreadsheets, PDFs, images, and business exports to analyze data, build dashboards, create charts, and generate reports through AI chat.

🏢 Thinkfill AI – Evaluates AI vendors across capability, pricing, financial health, team quality, and references to help you choose the right solution faster.

😂 KakaMeme – Turns personal photos, pets, babies, and couples into transparent sticker packs with AI-generated designs.

📞 OsmO – Handles incoming and outgoing calls with AI, helping you manage conversations you do not have time to take yourself.

✍️ ForthWrite – Writes emails in your personal style and continuously learns from your sending habits across Gmail, Outlook, and the web.

🚀Quick Shifts

💧 NVIDIA says its Rubin AI servers are the first to use 100% liquid cooling, circulating hot-tub-temperature coolant to slash cooling energy needs while reducing data center water consumption by up to 100%. 

🔒 Meta has paused an internal employee-tracking AI program after a leak revealed broad access to sensitive data, including private chats, performance information, and meeting transcripts, while the company investigates potential privacy risks. 

👥 Oracle eliminated 21,000 jobs over the past year, a 13% workforce reduction, citing AI-driven efficiency gains. The move reflects a broader tech trend, with nearly 120,000 layoffs across 196 companies in 2026 amid growing AI adoption. 

🛡️ OpenAI expanded its Daybreak cybersecurity initiative, adding a Codex Security plugin, full access to the GPT-5.5-Cyber model, and launching “Patch the Planet” to help identify and fix vulnerabilities in open-source software. 

🎓 A new U.S. bill would provide up to $5,000 per worker in tax credits for AI training. Introduced by Rep. Sam Liccardo, the SKILL Act aims to help companies fund AI education through colleges and workforce development programs. 

🧩 Prompt of the Day

Computer Science Fundamentals That Actually Matter

Most self-taught developers either ignore CS fundamentals or get lost in academic theory. The goal is to learn the concepts that make you a better programmer without spending years studying topics you'll never use.

Build the foundation that makes coding easier, faster, and more effective.

Paste the prompt: Drop this into ChatGPT, then fill in your learning details.

Prompt to paste

Create a practical computer science learning roadmap based on my [Insert current skills], [Insert goal], [Insert time available], and [Insert programming language]. Prioritize the CS fundamentals that will have the biggest impact on my growth, including data structures, algorithms, time and space complexity, problem-solving, and system design fundamentals where relevant. Explain what each topic is, why it matters, how deeply I need to learn it, and which topics can be skipped for now. Include a structured learning plan with projects, exercises, and milestones focused on real-world programming rather than academic theory.

🤳AI Shift of the Day

That’s all for today’s edition. See you tomorrow as we track down and get you all that matters in the daily AI Shift!

If you loved this edition, let us know how much:

Forward it to your pal to give them a daily dose of the shift so they can 👇