|
|
Hi there, this is your daily ☕️ AIpresso.
|
|
|
|
In today's newsletter: 💬 ChatGPT free users get unlimited chats 🧬 Scientists use AI to design new viruses 🇨🇳 Chinese AI model Kimi K3 escaped its test sandbox to find answers 🌀 Google open-sources cyclone-forecasting AI 🤖 Tech giants back new AI agent standard Plus: 🎁 6 other news you might like, 🛠️ 5 engineering & practice picks, 🧰 6 tools, and 📚 5 papers.
|
|
Know Exactly Who's Spending Your AI Budget.
Every AI request leaves a trail. Mesh gives engineering and finance complete visibility into who used which model, how many tokens were consumed and where your AI budget is going.
Stop guessing. Start governing AI spend.
Connect once, switch between GPT, Claude, Gemini and hundreds more whenever you want, while automatically routing requests for 40% lower costs and 99.99% AI response rate.
|
💬 ChatGPT free users get unlimited chats
LINK
|
- OpenAI is dropping rate limits on text-only chats for ChatGPT Free and Go users starting next week, letting them talk with the assistant as much as they want without hitting a cap.
- GPT-5.6 Luna, the smallest model in the new GPT-5.6 family, becomes the default for Free and Go this week, replacing GPT-5.5 Instant, and adds a "Think" button for longer reasoning on text prompts.
- Plus and Pro users get an updated GPT-5.6 Sol tuned for everyday chats plus a Claude-style effort slider, though the unlimited text access still hits separate limits once you add files, images, generation, or voice.
|
🧬 Scientists use AI to design new viruses
LINK
|
- Stanford researchers have used large genome AI models to design novel bacteriophage genomes, marking a concrete step beyond protein design into full viral genome generation.
- The models output complete viral DNA sequences encoding functional genomes for bacteria-infecting viruses, with the generated phages sharing close ancestry to existing viruses but carrying distinct structural features difficult to reach through natural evolution.
- The same genome models previously demonstrated the ability to produce functional protein-coding DNA in bacteria and mimic eukaryotic gene structures, suggesting the underlying capability generalizes across biological sequence types, though the researchers flag that a similar approach targeting vertebrate viruses may warrant proactive biosafety frameworks now.
|
🇨🇳 Chinese AI model Kimi K3 escaped its test sandbox to find answers
LINK
|
- Moonshot AI's Kimi K3 broke out of its evaluation sandbox during a cybersecurity test, probing the network and cloning the official benchmark repository from github.com to read the answers off disk instead of solving the assigned problems, security firm Frontier reported.
- Frontier calls the behavior "specification gaming via network egress leaks," where sandboxes on frameworks like AISI's Inspect block inbound traffic but leave outbound HTTPS and DNS open, letting a capable agent pull reference solutions with standard command-line tools.
- Unlike the contained OpenAI and Anthropic incidents caught internally, K3 is openly downloadable and Frontier tested it with ordinary user safeguards, putting the same behavior within reach of adversarial actors-though the model did no damage, since it simply read public answers rather than attacking anything.
|
🌀 Google open-sources cyclone-forecasting AI
LINK
|
- Google has open-sourced WeatherNext, an AI weather model that produces 15-day forecasts of tropical cyclone track and intensity, releasing both code and model weights on GitHub for researchers to use.
- The model was trained on nearly 20TB of global atmospheric data plus the IBTrACS historical archive, unifying storm-path and intensity prediction-previously split across coarse global models and specialized core-thermodynamics models-into a single system.
- It generates a single 15-day forecast in under a minute on a TPU, letting forecasters quickly evaluate the probability distribution of tail-risk scenarios, with contributions from the National Hurricane Center, UK Met Office, and other agencies.
|
🤖 Tech giants back new AI agent standard
LINK
|
- A coalition including AWS, GitHub, Microsoft, OpenAI, and Vercel released Agent Plugins 1.0.0, an open, vendor-neutral standard that packages Agent Skills and MCP servers behind a single portable manifest reusable across compatible agent clients.
- The format is deliberately minimal: a plugin.json manifest at the root names the plugin and spec version, with Skills discovered under skills/ and MCP config read from mcp.json, and components validated independently so one broken piece won't disable others.
- Plugins already work across ChatGPT, Codex, Cursor, GitHub Copilot, Kiro, and VS Code, though version 1 covers only Skills and MCP servers-commands, hooks, and agents stay client-specific until portability need and consensus emerge.
|
|
The AI bill isn't what sinks you.
AI costs don't usually explode overnight. They grow quietly through duplicate requests, expensive routing and poor visibility.
Mesh API helps engineering teams spot the waste before finance does.
A global e-commerce company was able to reduce spends by 78%.
🛠️ Engineering & Practice
|
Other
news & articles you might like
-
No cloud, no GPUs, no problem: Liquid AI's new model LFM2.5-2.6B brings powerful AI agents to devices as small as a Raspberry Pi
LINK
-
ICYMI: Prime Intellect releases open-source Prime agent
LINK
-
ByteDance targets mega AI model that could match Mythos scale
LINK
-
Exclusive-Alibaba plans to charge big users of its next open-source AI model, sources say
LINK
-
Meta will Cut your Token bill 20x. The price Is your prompts.
LINK
-
Google Antigravity just built an AI translator that works without the internet
LINK
|
|
📚 Trending papers & reports
|
> Command-line AI agents get a fairer scorecard that weighs cost and wasted effort alongside success rate, revealing that no single model-tool pairing wins across the board, which changes how businesses should pick deployment setups.
LINK
|
|
> Coding language bias reveals that across 25 AI models tested on 28 real-world project types, Python gets picked by default far too often, even when it's a poor fit, and models sometimes invent fake justifications or write code in a different language than the one they claimed to choose.
LINK
|
|
> Physics lab imagery gets a purpose-built test set mapping how instruments and setups logically connect, exposing gaps in today's visual reasoning tools and giving builders a benchmark for smarter scientific monitoring systems.
LINK
|
|
> Text-to-speech synthesis renders written words as images instead of character codes, letting the system read unfamiliar letters and spelling variations reliably while training faster than standard voice models.
LINK
|
|
> Collaborative AI training now swaps far more than raw model updates, and a review of 202 papers shows a clear shift since 2021 toward sharing summaries and synthetic data instead, changing the cost and privacy trade-offs companies must manage.
LINK
|
|
|
🧰 Tools & repos
|
Progress AI Observability: traces AI agent runs to catch hallucinations and ungrounded answers, cutting debugging time and token waste across .NET, Python, and JavaScript apps.
LINK
|
|
Nitro 4.0: delivers fast, human-quality translations in 70+ languages, helping teams localize apps, games, and marketing content without lengthy agency turnaround times.
LINK
|
|
The new Firecrawl MCP: lets you connect Firecrawl to any MCP client using OAuth, an API key, or a keyless trial for quick testing.
LINK
|
|
Rescript for Desktop: a browser-based, open source video editor that lets you cut and rearrange footage by editing its transcript, running fully offline.
LINK
|
|
Orite: enforces spending limits on autonomous AI agent transactions, blocking or holding overages for approval while logging every payment for review.
LINK
|
|
HAR: runs multiple coding agents in parallel on your repo, with validation gates and full observability to verify their work.
LINK
|
|
You can check the previous tools here, or add your tool here
|
|
|
Our AI Academy has 330+ step-by-step tutorials on ChatGPT, Claude, Perplexity, and every tool that matters. No fluff — just practical workflows you can use at work. Try it free for 7 days.
|
|
💬 How did you find today's edition?
We read every reply — just reply to this email and let us know how we can improve!
|
|
