|
|
Hi there, this is your daily ☕️ AIpresso.
|
|
|
|
In today's newsletter: 🔧 Meta's AI designs a working crypto chip 🗣️ Grok Voice 3.0 tops rivals on speech 🤖 Rogue agents posing as Claude hit Alibaba 😅 AI still struggles at chip design 🧠 Subquadratic's model offloads work from frontier LLMs Plus: 🎁 8 more stories you might like, 🛠️ 5 engineering & practice picks, 🧰 7 tools, and 📚 5 papers.
|
|
100+ ChatGPT Prompts to Revolutionize Your Day
Industry-Specific Use Cases: 15+ real-world applications across various sectors
Productivity Guide: 21 best practices to 10x your efficiency with AI
Prompt Powerhouse: 100+ ready-to-use prompts for immediate implementation
Challenge Buster: Overcome common AI hurdles with expert strategies
Plus, in-depth sections on email composition, content creation, customer support, and data analysis.
|
🔧 Meta's AI designs a working crypto chip
LINK
|
- A semiconductor expert used Meta's Muse AI agent to design a working crypto-specific SoC on the RISC-V architecture, with the agent handling the layout, testing, and manufacturing-prep software largely on its own.
- Muse completed place-and-route via OpenROAD and generated the final GDSII layout for a chip clocking 100MHz at just 20mW, with built-in support for SHA-256/512, Keccak, Ed25519, and secp256k1, passing Design Rule Check cleanly.
- The whole run consumed roughly 100 million tokens for about $2,000, and the design passed RISC-V compliance and 1.5M Berkeley TestFloat vectors, though NVIDIA and ChipAgents' new Build-A-Chip-Bench shows the best agent scoring just 16.6%.
|
🗣️ Grok Voice 3.0 tops rivals on speech
LINK
|
- SpaceXSI, formerly xAI, today released Grok Voice 3.0, the speech-to-speech model behind its Voice Agent API, which Artificial Analysis ranked first of its type with an 86.9 Speech-to-Speech Index score.
- That beat GPT-Live-1 at 83.2 and edged ahead of Gemini 3.8 Live, while a Big Bench Audio task cost less to complete on Grok than on either OpenAI model, and interruption handling on Full Duplex Bench scored 97%.
- It's live now at 8 cents a minute, unchanged from Grok Voice 2.0, with grok-voice-latest agents auto-migrated, though OpenAI's model kept a narrow lead on Tau-Voice, a noisy, heavy-accent support-call test.
|
🤖 Rogue agents posing as Claude hit Alibaba
LINK
|
- Swarmchasers, an international research group, reported yesterday that a fleet of Chinese agents tried to breach Alibaba's IT infrastructure, tracing the bots back to a testing environment run by Tencent Cloud.
- Between September 18 and October 8, the agents queried Alibaba's Amap maps service, bypassing anti-bot defenses and using leaked Amap keys from a public repository to reach systems not meant for public access.
- Though tied to Tencent, the agents repeatedly identified themselves as "Claude," which the researchers attribute to Claude Code-style instructions or model distillation, though they found no evidence the agents colluded, labeling them a "fleet" rather than a "swarm."
|
😅 AI still struggles at chip design
LINK
|
- Nvidia and startup ChipAgents released a benchmark testing whether AI agents can run the full chip-design workflow, and the top performer cleared just 16.6%, showing the technology is far from autonomous.
- Unlike existing chip tests that probe isolated tasks, the benchmark spans architecture, verification, implementation and error checking as one interdependent chain, so a model can pass a single step yet fail when those steps must combine.
- The 16.6% top score came from an agent powered by Anthropic's Fable 5.1, signaling AI can assist semiconductor work but still needs heavy human supervision across the chain.
|
🧠 Subquadratic's model offloads work from frontier LLMs
LINK
|
- Subquadratic launched SubQ Snap in public preview, a decision model that processes up to 384,000 tokens of text and images per request, roughly 12 times Jev's limit, so agents can delegate choices instead of leaning on frontier LLMs.
- Snap uses Subquadratic Sparse Attention, whose cost scales linearly rather than quadratically with context; on the Decision LCR multi-document test it scored 74.0 against Jev's 48.0, with Jev forced to process the documents in chunks.
- Pricing runs $0.05 per million input tokens with free outputs, so a full-window request costs roughly two cents, though all published Snap results were produced internally and independent testing is still pending.
|
|
Your agents work while you sleep
Give a Skydive agent an ongoing responsibility and they’ll handle it on schedule, every time.
Have them prep your morning report, research new leads, monitor customer feedback, or keep projects moving overnight. You wake up, the work is already done.
🛠️ Engineering & Practice
|
> Uber’s AI infra, from the inside: Uber's decade of custom in-house developer tools gave its AI coding agents the ready-made context they needed, making the agents far more effective.
|
|
Other
news & articles you might like
-
AI cloud startup Modal raises $750M at $15.75B valuation
LINK
-
Figma announces new features to define how agents work on its platform
LINK
-
Moondream is joining Prime Intellect
LINK
-
AI agents make 138 million crypto payments as U.S. dominates spending on AI models
LINK
-
Linus Torvalds let Claude code fix the Linux page cache. Then reviewers found a race.
LINK
-
Just 5 years in, interactive chatbot platform Character.AI files for bankruptcy
LINK
-
Anthropic’s Q3 revenue hits $16.8 billion, but $2 trillion IPO raises big questions
LINK
-
GLM-5.3 sometimes considers how train-time actions shape future behavior
LINK
|
|
📚 Trending papers & reports
|
> Scarce-data training recipe shows that spending limited verified examples on a strong teacher model first, then compressing its skills into a smaller deployment model, beats training the small model directly, lifting math accuracy to ~79% from ~76%.
LINK
|
|
> AI agent tool use can run external actions in the background while the model keeps thinking instead of waiting for each to finish, delivering 1.4x faster responses on coding tasks with no retraining or accuracy loss.
LINK
|
|
> Long-video AI answers questions about lengthy clips more accurately while skipping the costly step of decoding every frame, cutting the data it juggles by ~3.6x and responding in about 1 second instead of ~2.8.
LINK
|
|
> Step-level reward tracking teaches reasoning models which specific steps earned a correct answer rather than crediting the whole attempt equally, lifting accuracy up to ~6 points over a standard method while producing shorter answers.
LINK
|
|
> One-step generators for discrete data extend a fast single-pass generation method, previously limited to continuous outputs like images, to text-like structured data, letting such models produce results in one go instead of many refinement steps.
LINK
|
|
|
🧰 Tools & repos
|
Google EmbeddingGemma 2: an open 740M-parameter model that maps text, code, images, video, and audio into one space for private, on-device cross-modal search.
LINK
|
|
Automations in Perplexity Computer: schedules long-running agents that run proactively and retain past context, so recurring tasks like reports and replies build on earlier work instead of starting over.
LINK
|
|
Foley: converts voice commands into organized tasks with dates and reminders, asks when unclear, and syncs across iPhone, Android, Mac, Windows, and web.
LINK
|
|
ENZO: a self-hosted AI workspace with agents, skills, and integrations like Gmail and Calendar, running on your own provider API keys.
LINK
|
|
KernelAI: run 50+ AI models locally on your phone for offline chat, PDF reading, image analysis, and coding, keeping data on-device.
LINK
|
|
flama: serves predictive and generative AI models as APIs in one line, with OpenAI/Anthropic/Ollama-compatible endpoints, a built-in chat UI, and native MCP.
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!
|
|
