Gemini Flash upgrade ⚡️, Meta AI cloud 🌐, ZCode 👨‍💻

The inference layer your coding agents have been waiting for (Sponsor)

Coding agents only move as fast as the inference behind them. FriendliAI is the Frontier Inference Cloud for agents — serving frontier open-weight models with the speed, reliability, and cost efficiency agents need in production.

Why teams run agents on FriendliAI (see how):

> Built for speed: #1 output speed for GLM-5.1 on OpenRouter.

> Frontier models: GLM-5.2, MiniMax-M3, Kimi-K2.7

> Reliable at scale: 99.99% uptime SLA.

> Already powering: Claude Code, Cursor, Kilo Code, Hermes Agents, Ollama, and more.

👉 Run your agent with the fastest inference layer

🚀Headlines & Launches

Google might be testing Gemini Flash upgrade on LM Arena (2 minute read)

Google is testing a new Gemini Flash upgrade on LM Arena, reportedly offering incremental improvements over the current Flash version. Despite no official confirmation, past Arena tests have often preceded public launches, drawing attention to potential new labels like "Gemini 3.6 Flash" or "Gemini 4 Flash." The Flash tier attracts interest due to its importance for everyday users and cost-conscious developers, handling most free and pay-as-you-go traffic faster than pricier Pro models.

Meta Is Planning a Cloud Business to Sell AI Computing Power (3 minute read)

Meta Platforms is developing an internal cloud infrastructure initiative to sell access to its surplus AI computing power and hosted models to external developers. This strategic pivot aims to generate a new revenue stream from the company's massive multi-billion-dollar investments in data centers and chips, directly challenging dominant hyperscalers like Amazon Web Services, Microsoft Azure, and Google Cloud.

🧠Deep Dives & Analysis

Continual Harness: An Efficient Self-Improving Agent on ARC-AGI-3 (3 minute read)

The ARC-AGI-3 benchmark tests agents' ability to form and update an internal world model through ongoing learning. Continual Harness aims to efficiently enhance agents' self-improvement capabilities on this challenging IQ test.

Autoresearch: The feedback loop behind self-improving agents (11 minute read)

Autoresearch involves building an 'outer loop' where agents help maintain and improve the primary system using feedback signals, evals, and human input to make progress over time. One of the startups building infrastructure for these self-improving systems is Introspection. This post features an interview with Introspection's co-founder and CEO, Roland Gavrilescu. The discussion covers the shift from agent harnesses to feedback loops, the role of the open-source Pi framework, and why autonomous software factories must first learn from humans.

The Winning Essays for the Big Questions About AI (23 minute read)

Dwarkesh Patel posted some big questions about AI a couple of months ago and received 600 essay submissions for the contest. The three winners of the competition were Jassi Pannu, an Assistant Professor at Johns Hopkins University; Ege Erdil, the co-founder of Mechanize; and Michael Li, a Master of Public Policy candidate at Harvard Kennedy School. The full essays are available in the post.

👨‍💻Engineering & Research

ZCode (1 minute read)

ZCode is now available on macOS, Windows, and Linux. It combines the best AI agents with existing tools so developers can plan, code, review, and deploy without friction. GLM-5.2 is tuned for ZCode, making agentic coding faster and steadier. GLM Coding Plan subscribers now have 1.5x usage quota in ZCode.

Learning to Replicate Expert Judgment in Financial Tasks (14 minute read)

Frontier models struggle with relatively simple financial tasks. Custom models fine-tuned on high-quality proprietary datasets labeled by expert investors perform better. They are also substantially cheaper. Models of the future will probably feature differentiated intelligence - organizations will have custom models tuned to their specific needs that outperform frontier models.

PorTAL: Portable Task Adapters for LLMs (3 minute read)

PorTAL is an innovative architecture designed to decouple task fine-tuning from specific base model weights. As the accelerating cadence of foundation model releases creates a compounding financial and engineering burden to re-tune capabilities from scratch, PorTAL offers a way to pay for task adaptation once and amortize it across all future models.

🎁Miscellaneous

TLDR is hiring a curator for TLDR AI! (TLDR Curator, ~5 hrs/week)

Over 1M subscribers read TLDR AI to stay on top of the latest in AI models, research, engineering, and more. If you work in AI and want to help curate it, send your LinkedIn or resume to ai@tldr.tech!

A New Look at AI's Impact on Jobs: Firm-Level AI Spending and Workforce Adjustment (4 minute read)

A joint study by Ramp and Revelio Labs analyzed the relationship between firm-level generative AI investments and employment outcomes across over 21,000 companies in the US. The research reveals that companies with high-intensity AI spending grew their overall headcount by 10.2 percent and entry-level positions by 12 percent over the two years following adoption.

Redeploying Fable 5 (18 minute read)

Claude Fable 5 and Mythos 5 have been redeployed. Fable 5 will be included for up to 50% of weekly usage limits until July 7, after which it will be available via usage credits. Mythos 5 has been restored for some organizations in the US. Anthropic is continuing to coordinate with the US government to expand Mythos access to the broader set of domestic and international partners in the Glasswing program.