Sunday, June 7, 2026

What happened: Perplexity released a new "Search as Code" architecture that allows AI models to write their own search routines in Python instead of calling fixed APIs.

Key details:

• Models generate custom search pipelines with built-in filtering and deduplication logic running inside a sandbox environment
• The system beats OpenAI and Anthropic on key search benchmarks while reducing token costs by up to 85 percent
• Replaces rigid API-based search with flexible, agent-driven search logic
• Enables more granular control over information retrieval without fixed response formats

Why it matters: This approach challenges the conventional API-centric design for AI search, demonstrating that giving models control over search logic can simultaneously improve quality, reduce costs, and increase flexibility for downstream applications.

Practical takeaway: Evaluate Perplexity's Search as Code if your application requires efficient, high-quality web search with lower token consumption than current alternatives.

What happened: Meta announced two major AI product expansions: a paid AI agent service called "Hatch" priced up to $200 per month, and an AI-generated news feed feature in its standalone Meta AI app.

Key details:

• Hatch accepts natural language requests and autonomously builds working tools, schedules appointments, and sends emails
• Represents Meta's first major paid AI product offering and a strategy to monetize AI beyond advertising
• CEO Mark Zuckerberg framed Hatch as a way to refinance the company's massive AI infrastructure investments
• Meta AI app's "For You" section now auto-populates with AI-generated clickbait-style content including AI-generated images, text, and topics
• The AI-generated news feed mirrors decades-long patterns of algorithmic clickbait on Facebook, now powered by generative AI

Why it matters: Meta's move into premium AI services signals confidence in agent capabilities and reveals a critical business need: finding revenue models to offset growing AI infrastructure costs. Simultaneously, the AI-generated news feed represents a troubling expansion of synthetic content into core feed experiences, potentially amplifying misinformation at scale.

Practical takeaway: If you manage content strategy on Meta platforms, prepare for increased competition from AI-generated clickbait, and for enterprise customers, evaluate whether Hatch's $200/month pricing justifies its automation capabilities compared to traditional automation tools.

What happened: Sakana AI, a Japanese startup co-founded by Transformer architecture co-author Llion Jones, launched a dedicated research lab focused on recursive self-improvement (RSI)—AI systems that iteratively improve themselves.

Key details:

• Positions recursive self-improvement as an alternative path to competing in the AI capability race without participating in raw compute scaling
• Differs from frontier lab strategies (OpenAI, Anthropic, Google) that compete primarily on compute scale and training investment
• Anthropic has previously warned about control and safety risks associated with recursive self-improvement technology
• Reflects emerging debate within the industry about sustainability and viability of continued compute-driven competition

Why it matters: If recursive self-improvement can deliver capability gains more efficiently than compute scaling, it could reshape competitive dynamics and make frontier AI research more accessible to smaller labs. However, Anthropic's safety warnings indicate this approach carries distinct risks that may require new safety frameworks.

Practical takeaway: Watch Sakana AI's research output on recursive self-improvement—if successful, it could demonstrate a compute-efficient path to advanced AI capabilities and force reconsideration of infrastructure-heavy strategies.

What happened: Anthropic hired Clive Chan, who was the second hardware employee in OpenAI's custom chip program, while separately xAI was revealed to have trained its coding models on Anthropic's Claude outputs for months—continuing via private accounts and the Blackbox AI service even after Anthropic cut off access.

Key details:

• Clive Chan brings experience from Tesla's Autopilot ASIC development and the OpenAI-Broadcom partnership
• Anthropic is reportedly considering developing its own AI chips, signaling a hardware-first strategy as both companies prepare IPO filings
• xAI trained coding models on Claude outputs without permission and circumvented access cutoffs using private accounts and third-party services
• xAI's pretraining team shrank to fewer than five people with several leads departing
• The compute Musk purchased for xAI is now being rented to Anthropic and Google instead of powering xAI's own models

Why it matters: The talent poach reflects accelerating competition as both Anthropic and OpenAI race toward IPOs and recognize custom hardware as critical to scaling competitive advantages. Simultaneously, xAI's data practices reveal how competitive desperation can drive violations of training data boundaries, raising questions about industry norms around model-generated training data.

Practical takeaway: Monitor Anthropic's custom chip roadmap and be cautious about assuming your model outputs are protected from unauthorized training use by competitors.

What happened: Researchers published findings explaining why small language models fail at rare tasks: frequent tasks in training data continuously overwrite learning for uncommon skills.

Key details:

• Study examined models ranging from 4 million to 4 billion parameters
• The mechanism identified is task interference—common tasks overwrite rare task knowledge during training
• Proposed practical fix: increase the frequency of target rare tasks in training data instead of scaling up model size
• Implies that smaller models with properly balanced training data can match larger models on specific rare tasks

Why it matters: This finding challenges the conventional assumption that scaling is the solution to capability gaps and offers a more efficient alternative: targeted training data curation. For resource-constrained teams building specialized models, this suggests training data composition matters more than raw parameter count for achieving targeted performance.

Practical takeaway: When fine-tuning or training smaller models for specific rare tasks, prioritize increasing that task's representation in your training data rather than defaulting to larger base models.

What happened: OpenAI introduced Lockdown Mode for ChatGPT, a security setting that disables web access, Deep Research, and Agent Mode to reduce exposure to prompt injection attacks.

Key details:

• Lockdown Mode blocks the final step in an exfiltration chain but does not fully prevent prompt injection attacks
• Disables three capabilities simultaneously: web access, Deep Research feature, and autonomous Agent Mode
• Targets organizations and users handling sensitive data where prompt injection poses extraction risk
• Designed as a mitigation rather than a complete solution; prompt injection remains an unsolved problem at the protocol level

Why it matters: As AI systems gain access to external tools and data, prompt injection attacks have become a critical vector for data theft. This feature acknowledges the vulnerability while providing a practical defense for high-risk use cases, though it requires accepting significant capability restrictions.

Practical takeaway: Enable Lockdown Mode when using ChatGPT with sensitive proprietary or confidential information, but recognize it as damage control rather than a comprehensive security solution.