TLDR AI 2025-06-18

If slow QA processes bottleneck you or your software engineering team and you're releasing slower because of it — you need to check out QA Wolf.

They get engineering teams to 80% automated end-to-end test coverage and help them ship 5x faster by reducing QA cycles from hours to minutes.

QA Wolf takes testing off your plate. They can get you:

• Unlimited parallel test runs
• 24-hour maintenance and on-demand test creation
• Human-verified bug reports sent directly to your team
• Zero flakes, guaranteed

The result? Drata's team of 80+ engineers achieved 4x more test cases and 86% faster QA cycles.

No flakes, no delays, just better QA — that's QA Wolf.

Schedule a demo to learn more

Anysphere has introduced a fixed‑price Ultra tier that offers far higher compute than Pro, made possible by long‑term deals with major model vendors.

Google released Gemini 2.5 Pro and Flash to general availability, debuted Flash‑Lite in preview, and introduced controllable “thinking” budgets that improve reasoning accuracy while giving developers flexibility.

In a podcast, the OpenAI CEO confirmed Meta dangled nine‑figure packages to lure researchers to its superintelligence team but said virtually none defected, taking the opportunity to jab at Meta's recruiting push.

This career guide identifies skills that paradoxically become more valuable as automation increases, including AI deployment, leadership, and government relations. It advises knowledge workers to try to leapfrog traditional entry-level positions entirely, advocating for side projects and startup roles as AI hollows out established corporate hierarchies.

Key-value (KV) caches store intermediate attention computations during LLM inference to avoid redundant calculations. When generating "Time flies fast" token by token, the model normally recomputes attention for "Time" and "flies" at each step, but caching these values delivers 5x speedups. This tutorial progresses from modifying a 124M parameter GPT model with basic cache buffers and position tracking to production-ready optimizations like pre-allocated memory and sliding windows to address the linear memory growth that becomes prohibitive for long sequences.

This guide emphasizes starting with single agents before multi-agent systems, using manager patterns where one agent coordinates others via tool calls or decentralized handoffs for peer-to-peer task delegation. Key insights include implementing guardrails as layered defenses (LLM-based classifiers, regex filters, and moderation APIs), designing tools for messy long-horizon tasks, and building human-in-the-loop mechanisms triggered by failure thresholds or high-risk actions.

Delve ran 4 placements across 2 TLDR newsletters and brought in 66 leads, with a meaningful chunk from enterprise. The result: $1M in attributed pipeline and a 52x ROI.

The case study breaks down their strategy, results, and example ads.

Read the Delve case study.

TCANet layers multi‑scale convolutions, temporal compression, and stacked self‑attention to model motor‑imagery EEG.

Unlike chatbots or image generators, robots must operate in real time - delays between inputs and outputs have a tangible effect on performance. While vision-language-action (VLA) models have achieved promising results in open-world generalization, they can be slow to run. This article presents an algorithm called real-time chunking that enables real-time execution without discontinuities. It works on any diffusion- or flow-based VLA with no training-time changes.

MiniMax's 456B parameter model uses a hybrid mixture-of-experts architecture with "lightning attention" that processes 1 million token contexts (8x DeepSeek R1) while using 25% fewer FLOPs at 100K token generation lengths.

Briefings on AI risks revealed most UK parliamentarians lack in-depth AI knowledge and face capacity constraints that limit research on AI issues. The briefings were well-received, with one in three lawmakers publicly supporting AI risk mitigation campaigns. Effective outreach included persistent follow-ups and leveraging statements from notable AI authorities to convey the seriousness of AI-related extinction risks.

o3-pro seems to provide better answers than o3, but it has a significantly longer wait time. Using the API at scale seems prohibitively expensive, so users should instead run parallel queries using the chat interface. o3-pro serves the same niche as o3, so users thinking of using Opus may prefer to use Opus rather than, or in addition to, o3-pro. The 80% price cut on o3 seems more impactful than o3-pro - o3-pro is still largely a 'special cases only' model.

You'll probably like our flagship newsletter. It's all about tech, science, and programming.

Same quick format. Still free.

Subscribe now.

WhatsApp is testing an "AI Studio" feature that allows users to create custom chatbots without coding.

Anysphere, the developer of the popular artificial intelligence code editor Cursor, has been approached by investors about a deal that would double its valuation.

Gemini 2.5 Pro got into various situations which caused it to simulate panic - it made poor, hasty decisions when under stress.

Meta is co‑developing fashion‑forward AI eyewear with Prada, expanding beyond its Ray‑Ban tie‑up and hinting at a broader lineup that may soon include $360 Oakley glasses.

The AI startup needs fresh capital as it burns through $1 billion monthly, with Tuesday's commitment deadline for the debt sale testing investor appetite amid fierce competition for AI funding.