SunBrief#68: Gemini 3 Deep Think is now the world’s top-performing model

Welcome to the SunBrief

Today in SunBrief 🌞

• Transform Your Health with TruDiagnostic

• Gemini 3 Deep Think Gets Major Upgrade for Science, Research, and Engineering

• Google Warns Attackers Are Wiring AI Directly Into Live Cyberattacks

• Stock Updates

• Elon Musk says xAI departures were firings not resignations

• AI Highlights of the Week

• Too Important to Miss

Transform Your Health with TruDiagnostic

Take control of your health with the most comprehensive at-home blood test available. TruDiagnostic analyzes over 185 biomarkers from a single finger prick of blood, providing personalized recommendations for the most impactful changes you can make today.

Discover your biological age, understand the pace at which you’re aging, and get a detailed action plan tailored to your body. All from the comfort of home. [Newsletter Name] readers get 20% off with code NEWSLETTERS20

Gemini 3 Deep Think Gets Major Upgrade for Science, Research, and Engineering

Google updates its most specialized reasoning mode and expands access in the Gemini app and (select) API

Google has rolled out an upgraded Gemini 3 Deep Think designed to tackle complex, open-ended problems in science research and engineering, where data is incomplete and clear answers do not always exist

Key Points:

• Built for Research-Grade Reasoning: The update was developed with input from scientists and researchers to better handle complex, real-world research work not just clean benchmark-style problems.

• Wider Access Surfaces: Deep Think is available in the Gemini app for Google AI Ultra subscribers, and Google is also opening API access to select researchers, engineers, and enterprises via an early access pathway.

• Stronger Performance on Hard Benchmarks: Google highlights new peak results across rigorous evaluations, including Humanity’s Last Exam, ARC-AGI-2, and Codeforces, plus gold-medal-level performance claims for elite math competitions.

• Expanded Strength Beyond Math/Coding: The updated mode is positioned as strong across broader scientific domains like physics and chemistry, not only abstract reasoning.

• Practical Engineering Utility: Deep Think is pitched for applied workflows—like interpreting complex data, modeling systems through code, and even turning a sketch into a 3D-printable design.

Why It Matters:
This moves beyond chatbot features toward a deployable research assistant, bringing advanced reasoning into the Gemini app and API. If it performs well in real use, it could significantly reduce time spent on analysis, hypothesis testing, and complex debugging.

Google Warns Attackers Are Wiring AI Directly Into Live Cyberattacks

Threat actors move from “testing AI” to embedding it inside real-world malware and attack workflows

Google’s Threat Intelligence Group warns that attackers are no longer just testing generative AI. They are actively using it in live cyberattacks to write code, speed up intrusions, and probe defenses in real time.

Key Points:

• AI in the Attack Chain: Google observed malware families making direct API calls to Gemini during runtime, requesting generated code on-demand instead of shipping all logic inside the binary.

• Harder-to-Detect Malware: By shifting functionality outside static payloads, attackers can reduce reliance on fixed signatures and make traditional detection (static + some behavioral) less effective.

• Example Malware Use: A malware family Google identified as HONESTCUE reportedly used prompts to retrieve C# code that was then executed as part of the intrusion workflow.

• Model Extraction Attempts: GTIG also describes distillation/model extraction activity large volumes of structured prompts designed to infer model behavior and replicate proprietary capabilities without building a model from scratch.

• AI as Operator Support: Google says both state-aligned and financially motivated groups are using AI to enhance recon, vulnerability research, script development, and phishing content generation.

Why It Matters:
This marks a shift from AI merely speeding up attackers to becoming part of live attacks. On-demand code generation, stronger lures, and faster iteration shrink the time from idea to intrusion,n forcing defenders to adapt to an AI-driven threat landscape.

Stock Updates

Elon Musk says xAI departures were firings not resignations

Two more co-founders leave as Musk says the company “parted ways” during a speed-focused restructuring and is hiring aggressively

Elon Musk is responding to a fast wave of departures at xAI. Two more co-founders left this week, taking the total to six of the original twelve. Musk says this is about scaling the company and moving faster, not about individual performance, and he has hinted the latest exits were not voluntary.

Key Points:

• Six of Twelve Co-Founders Have Left: With two additional co-founders departing this week, half of the original founding group is now gone.

• Musk First Said “Fit,” Then Implied “Let Go”: At a Tuesday night all-hands meeting, he suggested some people are better for early-stage building than later stage scale. The next day on X, he went further and said a reorganization “required parting ways with some people.”

• He Framed It as a Speed Upgrade: Musk described the changes as a structural shift to improve the speed of execution as the company grows quickly.

• It was not just the founders: Reporting says at least 11 engineers have publicly posted about leaving in roughly a week (some departures happened earlier but were announced recently), which helped the “mass exodus” story spread online.

• Some Leavers Hint at a New Startup: Several departing staff said they want to start something new, often pointing to the idea that small teams with strong AI tools can build faster and with more freedom.

• Tough Timing Around Scrutiny and Big Moves: The exits come amid heightened scrutiny, including investigations into harmful and nonconsensual deepfake content tied to Grok and activity on X. French authorities reportedly raided X’s Paris offices on February 3, 2026. The timing also follows SpaceX’s announcement that it plans to acquire xAI.

Why It Matters:
Losing half your cofounders is unusual for a fast-growing AI lab. Calling it a necessary reorg looks like an attempt to get ahead of a narrative already spreading online. Even if operations continue at scale, repeated exits can hurt recruiting, morale, and external trust, which matters when competing with OpenAI, Anthropic, and Google for top talent.

AI Highlights of the Week

• OpenAI to Remove Access to Sycophancy-Prone GPT-4o Model

  OpenAI will stop access to GPT-4o starting Friday, retiring it alongside four other legacy ChatGPT models.

  The company says only 0.1% of users still select GPT-4o (about 800,000 weekly users), but the decision has triggered major backlash, with users citing close relationships with the model and ongoing debate around sycophancy and user safety.

• Anthropic Raises 30 Billion Dollars at 380 Billion Valuation

  Anthropic has raised 30 billion dollars in Series G funding, valuing the company at 380 billion post-money, led by GIC and Coatue.

  The capital will fuel frontier AI research, enterprise products, and infrastructure expansion, as Claude adoption accelerates across large enterprises and developers..

• OpenAI Unveils Codex Spark With 15x Faster Coding

  OpenAI has introduced GPT-5.3 Codex Spark, a new ultra-fast coding model built for real-time conversational programming, delivering up to 15x faster code generation than GPT 5.3 Codex.

  Access is currently limited to 200 dollar per month Pro users, with trade offs in complex reasoning depth as the model runs on Cerebras AI chips in an early preview tier.

• SeeDance 2.0 Sparks Surge in Hyper Realistic AI Video Buzz

  Interest in ByteDance SeeDance 2.0 has hit an all-time high, with early tests showing AI-generated videos that closely resemble real footage using only simple prompts.

  Viral clips featuring hyper realistic motion have drawn criticism from the Motion Picture Association, intensifying debate around copyright, deepfakes, and the shrinking gap between real and synthetic video.

Too Important to Miss

Z.ai Introduces GLM-5 for Agentic Systems Engineering

Waymo Deploys Next-Gen Ojai Robotaxis to Extend U.S. Lead

ByteDance Unveils Doubao 2.0 as Disney Raises Copyright Claims

Spotify Says Its Best Engineers Stopped Writing Code Thanks to AI

Last Week’s Poll Result

• Which AI do you prefer overall right now?

  Depends on the task → 42.86%

  Claude Opus 4.5 → 28.57%

  GPT-5.3-Coder → 28.57%

• Should government contractors face stricter cybersecurity requirements than private companies?

  Yes, they handle extremely sensitive data → 69.57%

  Maybe, but standards should be universal → 26.09%

  No, current regulations are sufficient → 4.35%

• Should governments invest more in AI-driven biotech for rare diseases?

  Yes, public funding is essential → 66.67%

  Maybe, alongside private investment → 33.33%

Feedback

Know someone who may be interested?