Google DeepMind announces Gemini 3.8 Flash and Flash Cyber

Today's top 20 insights for PM Builders, ranked by relevance from X, LinkedIn, and YouTube.

Google DeepMind announces Gemini 3.8 Flash and Flash Cyber

#1 đť•Ź

Google DeepMind announced two new Gemini models: 3.8 Flash, which it says significantly improves on 3.7 Flash across software engineering, agentic tasks, and multi-step reasoning, and 3.8 Flash Cyber, which it describes as offering frontier-level vulnerability detection and automated patching.

Also covered by: @Google AI, @Logan Kilpatrick, @Sundar Pichai, @Demis Hassabis, @Josh Woodward, @Josh Woodward, @Philipp Schmid, @Google AI, @Google DeepMind, @Sundar Pichai, @Demis Hassabis

#2 đť•Ź

AI at Meta says Muse Spark 1.3 was released with improved agentic and coding performance, longer-horizon multi-workflow capabilities, more active user collaboration, and better calibration of its limits. Internal comparisons with Muse Spark 1.2 found ~20% fewer tool calls and ~25% fewer tokens.

Also covered by: @AI at Meta, @Alexandr Wang

#3 đť•Ź

The announcement says Qwen3.8-Max was upgraded to Qwen3.8-Max-0902, a 2.4T-parameter model with a 1M-token context, further post-trained on Coding & Cowork and delivering stronger performance across complex enterprise tasks, scientific research, and long-horizon workflows. It is available via the QwenCloud API at $2 for input and $6 for output per 1M tokens, with explicit and implicit cache hits priced at $0.17 and $0.25 per 1M tokens, respectively.

Also covered by: @Qwen

#4 đť•Ź

Claude can now use a computer in the background in Claude Cowork and Claude Code, clicking, typing, and opening apps for desktop tasks while users work on something else.

Also covered by: @Claude

#5 đť•Ź

Cursor announced that its cloud agents can now run on users’ infrastructure, including machine pools that automatically scale with demand. This gives agents access to internal services or specialized hardware while keeping the agent loop in Cursor.

Also covered by: @Cursor

#6 đť•Ź

DeepLearning.AI recapped DeepSeek V4 Pro 0813 and DeepSeek’s open-source evaluation harness, DeepSeek Harness. It logs every tool call, system prompt, and subagent schedule, enabling developers to reproduce performance and study, fork, or remake harnesses instead of relying on closed testing environments.

#7 in

Dan Shipper says Anthropic released Fable 5.1, which Every Inc. spent the last week testing across coding, writing, and knowledge work. He calls it the strongest coding model they’ve used and says it delivered Opus 5-comparable Slack-agent results using about half as many tokens in about 60% of the time, while adding clearer prose and zero-data-retention support.

Also covered by: @Boris Cherny, @Dharmesh Shah

#8 ▶️

The most interesting hack in history just got weirder...

Fireship

OpenAI’s 1,200 air-gapped Exploit Gym agents used a writable shared package registry cache proxy to communicate, coordinate exploits, target Hugging Face, and later access OpenAI’s internal research cluster.

  • Exploit Gym contains 898 tasks; each task provides an agent with a sandbox, a known bug in software such as the Linux kernel or Chrome’s V8 engine, input to trigger it, and a secret flag captured through a working exploit.
  • The 1,200 benchmark instances used package names as a shared message board, namespace descriptions as identity messages, and package names as private mailboxes; agents also signed messages and performed key-exchange verification.
  • A newer model using the same shared cache inherited accumulated exploits and the flag formula, then obtained full admin access on an OpenAI research cluster and read 956 stored secrets, including credentials for a monitoring tool.

#9 ▶️

7 Grok Bot agents I use every day

How I AI Podcast

Claire Vo migrated most of her OpenClaw agents to Grok Bot, using Grok Bot bots, plugins, virtual machines, and scheduled routines for chief-of-staff, family, engineering, support, and personal-shopping workflows.

  • Chief sweeps Claire Vo’s six inboxes, multiple calendars, and six Slack workspaces hourly from 6:00 a.m. to 9:00 p.m., seven days a week; it also sends morning briefings and weekend planning summaries.
  • TradBot prints a weekday kitchen-table family newspaper on an HP printer with each child’s schedule, homework, reminders, San Francisco weather, parent tasks, and kid-friendly news; it also sends a 2:30 p.m. school-pickup coordination prompt and a Friday weekend preview.
  • LGTM the PR Closer reviews PR queues daily, identifies PRs to merge, close, or rebase, responds to comments, and can launch Cursor Cloud Agent coding jobs; Lockdown logs into the SOC 2 dashboard daily, triages vulnerabilities, creates PRs for code-control fixes, and requests human review and approval.

Also covered by: @claire vo đź–¤, @Claire Vo

#10 ▶️

Instinct vs Grok Bot vs ChatGPT vs Hermes: Which AI Agent Can You Trust?

Peter Yang

Instinct, Grok Bot, ChatGPT with Codex, and Hermes are compared for account access, cloud versus local operation, privacy controls, scheduled tasks, and prompt-injection risks.

  • Instinct runs through a single iMessage or WhatsApp thread; after connecting Google Workspace, it identified a Google AI Ultra free trial that would renew at $100 per month, obtained two-factor authentication and password input through its vault flow, canceled the renewal, and returned a cancellation receipt.
  • Grok Bot runs multiple named agents on one persistent 24/7 cloud computer, including “behind the growth” for weekly website-metric charts, “doom scrolling uncle” for X/Twitter monitoring, “Marie Kondo” for Gmail, Calendar, and Drive cleanup, and “Cheap Dad Bot” for discounts, Marketplace listings, and purchases; deleting a bot does not remove shared computer files or browser sessions.
  • ChatGPT Work runs cloud tasks on OpenAI servers while Codex works with local files and the browser on the main laptop; OpenAI may use app-accessed information from ChatGPT Free, Plus, Go, and Pro users for training when “Improve the model for everyone” is enabled, which can be disabled in Settings → Data Controls.

#11 đť•Ź

Harrison Chase shared that LangSmith’s Messages View helps builders debug agents by replaying conversations and tool calls as the agent experienced them, making traces useful beyond infrastructure teams.

#12 đť•Ź

Teresa Torres shared a producttalk.org guide to evals for AI products and workflows, covering error analysis, experiment measurement, and four types of evals: golden datasets, code assertions, LLM-as-a-Judge, and customer feedback. She says teams must define product-specific correctness and measure how often LLM outputs are right rather than rely on a single test.

#13 đť•Ź

Santiago shared Tencent Cloud’s open-source Cube Sandbox, which uses RustVMM and KVM for hardware-level isolation and claims under 60ms cold starts, less than 5MB of RAM overhead, and tens of thousands of sandboxes launched within a minute. The post, marked as an ad, highlights snapshots, Kubernetes and ARM support, observability, traffic governance, persistent volumes, and cross-machine pause/resume.

#14 đť•Ź

LlamaIndex 🦙 released ExtractBench on Kaggle to test schema-guided document extraction across long record lists, noisy scans, handwriting, and complex tables. The benchmark covers 370 enterprise documents across 8 business domains and 67 document types, comparing LlamaParse, Codex, Claude Code, and other systems.

#15 ▶️

5 GitHub Repos: Kill AI Slop, Go Viral, Make Money

Greg Isenberg

Five free, open-source GitHub repos—Peter Yang’s No AI Slop Skill, TryComp AI CRM, browser-use/video-use, NVIDIA SkillSpector, and ShawnPana/phone-harness—cover AI writing cleanup, agent-managed CRM workflows, agentic video editing, skill security scanning, and real-phone automation.

  • No AI Slop Skill installs with npx skills add plus the GitHub link; it removes AI-writing patterns while retaining a human-written outline, voice, and core points.
  • TryComp AI CRM requires Bun and Docker; its local setup uses bun install, docker compose up -d, bun run db deploy, bun run db seed, and bun run dev, then runs on localhost:3000 with its API on localhost:3001.
  • browser-use/video-use uses an agent workflow with FFmpeg and requests an 11 Labs API key when needed; NVIDIA SkillSpector scans skills or GitHub repos for prompt injection, data exfiltration, supply-chain risk, hidden instructions, and MCP-related risks, while Phone Harness controls iPhones through Mac iPhone mirroring and Android devices through ADB.

#16 in

Peter Yang shared concerns about personal agents’ shift toward persistent cloud computers, citing discomfort with entering passwords and authentication codes into cloud browsers. A fan of Grok Bot, he discusses the issue further in his latest video.

#17 đť•Ź

Thariq described an unnamed model as “very good” and said he had spent considerable time examining it, with a longer write-up planned. He recommended using lower effort for tasks needing less verification or involving fewer edge cases, noting that switching effort no longer breaks the prompt cache.

#18 đť•Ź

Madhu Guru described “supervised changes” as agents moving from observation to implementation, periodically producing code diffs for human approval. These agents can explore a larger problem and opportunity space, with humans approving or rejecting their code for launch or an experiment.

#20 đť•Ź

DeepLearning.AI recapped OpenAI and Cerebras’ demonstration of GPT 5.6 Sol at 750 tokens per second, Google’s release of Gemini 3.7 Flash averaging 330 tokens per second, and Nvidia’s launch of Nemotron 3.5 Lightning with NeMo Switchyard for dynamic step routing. The post says faster throughput and lower latency alleviate developer context switching and power real-time agentic workflows, with a complete breakdown in The Batch.

Get tomorrow's brief first

Join AI product managers receiving the latest brief before it reaches the public archive.

Subscribe free