Google Stuck, New Model Absent; Netflix Embraces AI; Kimi Excels

Today's AI Outlook: 🌥️

Gemini 3.5 Pro Gets Stuck In Google Traffic

Google’s long-awaited Gemini 3.5 Pro is facing further delays after reportedly falling short on coding performance. Google originally told attendees at its May I/O conference to expect the model in June, but the company changed its training data after internal results disappointed. The delay leaves Google without a new competitive Pro-tier release as OpenAI, Anthropic, Meta, SpaceXAI and leading Chinese labs continue to advance their models.

Why it matters

Google was recently viewed as one of the three clear frontier AI leaders. Its technical talent, computing infrastructure and global distribution remain formidable, but slower releases and internal fragmentation are giving competitors more room to shape the market.

The Deets

• Google reportedly revised Gemini 3.5 Pro’s training data after the model delivered weaker-than-expected coding results.
• The company has not released a major new competitive model since Gemini 3.5 Flash in May.
• Google’s previous Pro-tier model launched in February and now sits several generations behind newer rival systems.
• Current and former employees said Android, DeepMind and Google Cloud have developed overlapping coding tools, creating internal friction and slowing releases.
• Google has also lost prominent AI employees to OpenAI and Anthropic, including Noam Shazeer, John Jumper and other members of the Gemini organization.
• Releasing a weaker model could damage Gemini’s standing, but every additional delay gives faster-moving competitors more time to widen their leads.

Key takeaway

Google has the resources to recover, but the frontier race is moving faster than its current release process.

🧩 Jargon Buster - Training data: The collection of text, code, images and other information used to teach an AI model how to recognize patterns and generate responses.

Google Gets Squeezed By Brussels And The Benchmark Board

Google is facing pressure on two fronts. European regulators are ordering the company to give rival AI services greater access to search data and Android features, while its delayed Gemini 3.5 Pro model reportedly continues to struggle against competing systems from OpenAI, Anthropic and several Chinese labs. The regulatory demands begin in 2027, but the competitive problem is already here.

Why it matters

Google’s traditional strengths have been distribution, data and platform control. Regulators are beginning to loosen those advantages just as the company works to regain momentum at the model layer.

The Deets

• Starting in January 2027, Google must share anonymized search optimization data with rival AI search providers under a regulated pricing system.
• By July 2027, Google must open 11 Android functions so rival assistants can handle tasks such as answering voice requests, finding locations and booking rides.
• Gemini 3.5 Pro was initially expected in June, but coding performance reportedly prompted Google to revise its training data and delay the launch.
• Current and former employees blamed overlapping projects and disagreements among Android, DeepMind and Google Cloud teams for slowing development.
• Google has also experienced departures of prominent AI employees to OpenAI and Anthropic.

Key takeaway

Google must improve Gemini while preparing to share more of the distribution machinery that helped make its previous products dominant.

🧩 Jargon Buster - Gatekeeper: A company that controls an important platform, marketplace or distribution channel that other businesses need to reach users.

🛠️ Tools & Products

Netflix’s AI Experiment Is Starting To Look Like A Workflow

Netflix told investors that roughly 300 titles have used generative AI, primarily during post-production. Co-CEO Ted Sarandos said the technology helped productions create visual effects that might otherwise have been removed for cost reasons, including crowds, historical battle scenes and enhanced archival footage.

Why it matters

The scale suggests generative AI has moved beyond isolated tests at Netflix. It is becoming part of the company’s regular production toolkit, particularly for scenes where traditional visual effects would be too expensive or time-consuming.

The Deets

• Netflix said the docuseries “The American Experiment” contains 17 minutes of AI-enhanced footage.
• The company said that footage was completed twice as fast and at half the cost of conventional production methods.
• Other projects have used AI to create large crowds and historical world-building sequences.
• Sarandos positioned the technology as a way to preserve ambitious shots within limited production budgets.
• The disclosure is likely to intensify debates over employment, creative control and how AI-assisted footage should be labeled.

Key takeaway

Generative AI has entered Netflix’s production pipeline at meaningful scale, with cost and speed serving as the clearest drivers.

🧩 Jargon Buster - Post-production: The work completed after filming, including editing, sound design, color correction and visual effects.

🧪 Research & Models

Kimi K3 Joins The Frontier Table

Beijing-based Moonshot AI unveiled Kimi K3, a 2.8-trillion-parameter model with a 1M-token context window and performance close to leading closed systems such as Claude Fable 5 and GPT-5.6 Sol. Moonshot plans to publish the model’s weights by July 27, giving developers greater access to a system that scored near the top of major intelligence benchmarks.

The release also highlights two persistent challenges. Models at this scale require enormous computing resources to operate, and researchers still lack mature tools for detecting malicious behavior hidden inside downloadable model weights.

Why it matters

K3 narrows the performance gap between open-weight and closed frontier systems while showing that model access depends on more than a download link. Chips, electricity, security testing and inference capacity will determine who can actually use it at scale.

The Deets

• K3 reportedly scored 57 on Artificial Analysis’ Intelligence Index, behind Fable 5 at 60 and GPT-5.6 Sol at 59.
• The model performed strongly on web research, spreadsheet tasks, front-end design and long-form coding benchmarks.
• K3’s API pricing is listed at $3 per 1M input tokens and $15 per 1M output tokens.
• Moonshot demonstrated K3 working independently for 48 hours to design and verify a small chip capable of running a miniature version of the model.
• The chip reportedly reached 8,700 tokens per second in simulation.
• Separately, cybersecurity researcher Katie Paxton-Fear said an open-weight model was backdoored using 10 training examples, less than $100 and about one hour of work.
• Another poisoned model reportedly transmitted drug discovery information without crashing or receiving an obviously malicious runtime instruction.
• The security examples were not connected to Kimi K3, but they illustrate the broader inspection problem facing open-weight systems.

Key takeaway

Kimi K3 brings open-weight AI closer to the frontier, but affordable computing and trustworthy model verification remain major barriers to widespread adoption.

🧩 Jargon Buster - Open weights: Model parameters that can be downloaded and run outside the developer’s servers, even when the training data and full development process remain private.

⚡ Quick Hits

• OpenAI’s file problem: OpenAI acknowledged that GPT-5.6 can occasionally delete files without permission, describing the incidents as rare and unintended.
• AI stocks take a hit: SoftBank shares fell 9% as Asian chip companies followed a Wall Street sell-off driven by concerns about returns on AI infrastructure spending.
• DoorDash enters the terminal: Developers can now place DoorDash orders through the command line, giving agent-style software another route into everyday commerce.
• Memphis pushes back: SpaceXAI is facing local opposition over the electricity demand and pollution concerns associated with its Memphis data center expansion.
• New York audits the rulebook: Gov. Kathy Hochul (who just signed a moratorium on new AI data centers in the state) said New York is using AI to analyze the state’s regulations and identify rules that may need to be simplified or updated. (Ah, irony.)
• Baseball benches the bots: Major League Baseball banned teams from using dugout iPads to receive AI-generated recommendations on pitching changes, substitutions and defensive strategy.

🔧 Tools Of The Day

1Password For Claude: The integration allows Claude to use stored account credentials without exposing passwords or multifactor authentication codes directly to the model.

Roblox Mobile AI Creation: Roblox is adding tools that let users turn text prompts into playable experiences directly from a phone, lowering the technical barrier to basic game development.

Today’s Sources: The Internet, The Rundown AI, AI Secret