Google launches Gemini 3.8 Live

Hi there, this is your daily ☕️ Devshot.

In today's Devshot:

🗣️ Google launches Gemini 3.8 Live

🧠 ChatGPT co-creator launches Jev model

🤖 Meta AI agents set up WhatsApp Business

🔍 Pinterest cuts search memory use

☕ Java 27 adds post-quantum cryptography

Plus: 🎁 6 other news you might like, 🧰 6 tools, and 📚 5 papers.

🗣️ Google launches Gemini 3.8 Live LINK

Google DeepMind shipped Gemini 3.8 Live and a 3.8 Live Extended Thinking variant, two audio models for building voice agents, now available to developers through the Gemini API and Google AI Studio.

Gemini 3.8 Live drives voice agents that make background API calls, handle visual input, and speak simultaneously across 97+ languages, while Extended Thinking tops the Artificial Analysis speech-to-speech leaderboard at ~83%, ahead of OpenAI's GPT-Live-1.

Pricing runs $0.005 per minute for audio input and $0.018 for output, putting an hour of conversation near $1.38 versus at least $3 on GPT-Live-1, though OpenAI's full-duplex model still sounds more natural.

🧠 ChatGPT co-creator launches Jev model LINK

TypeSafe, founded by ChatGPT co-inventor Diogo Almeida, launched Jev, a model that skips text generation to output type-safe structured values from unstructured input, aimed at automating deterministic decision logic inside production codebases.

Rather than autoregressively emitting tokens, Jev uses a parallel sampler that returns all structured values at once, constrained to predefined schemas so parsing pipelines and guardrail wrappers become unnecessary, handling selections across up to 255 options.

Internal tests clocked 70-500ms latencies versus seconds for conversational models, input priced at $0.042 per million tokens with output tokens free; early developer access opened today for waitlisted teams.

🤖 Meta AI agents set up WhatsApp Business LINK

Meta shipped a WhatsApp Business Tools MCP server that lets AI coding agents like Claude, Cursor, Codex, and ChatGPT set up and manage WhatsApp Business messaging for companies through natural-language instructions in chat.

The server handles onboarding tasks that previously spanned the Developer Console, Business Manager, and API references: creating a Business account, adding and verifying a phone number, registering for Cloud API access, and checking Terms of Service requirements.

Agents can also create or edit messaging templates, test messages and webhooks, and monitor Terms of Service status, payment methods, and Business Verification, while Meta's Social Technologies MCP server helps discover endpoints and troubleshoot errors.

🔍 Pinterest cuts search memory use LINK

Pinterest reworked its Manas search platform to serve tens of billions of embeddings using vector quantization, SSD-based indexing, and late-interaction retrieval, cutting memory footprints and yielding 20-30% serving cost savings across production workloads.

On a 100-million-embedding dataset, Product Quantization shrank HNSW indices by 74% and IVF indices by 93% at 70-80% recall, while Scalar Quantization cut them 59% and 75% respectively, holding recall above 90%.

SPANN with PQ hit 3x the QPS of DiskANN at a third the latency with a 5% recall drop, and across 5 billion embeddings saved over 40% CPU time versus full in-memory HNSW.

☕ Java 27 adds post-quantum cryptography LINK

Oracle shipped Java 27, folding in post-quantum cryptography through hybrid key exchange for TLS 1.3 that pairs classical algorithms with the module-lattice key encapsulation mechanism delivered back in JDK 24 under RFC 8446.

The security push spans several JEPs added since JDK 24, including quantum-resistant lattice-based key encapsulation and digital signatures, a Key Derivation Function API, and PEM encoding of keys, certificates, and revocation lists, now in third preview.

This non-LTS release also makes G1 the default garbage collector everywhere, advances primitive types in patterns and switch to a fifth preview, and bundles JavaFX 27's new Metal rendering pipeline on macOS.

Other news you might like

• Inside OpenAI’s agentic software factoryLINK

• Fedora Linux 45 Beta is here with big security and desktop changesLINK

• Give every teammate and agent the right level of access to your WorkersLINK

• Bolt is giving developers 50x more compute. But there’s a catch.LINK

• Every user, device and agent gets its own table in KeewanoDB, and queries never join across themLINK

• Kubernetes v1.37: Pod-Level Resource Managers graduated to BetaLINK

🧰 Trending tools

SKI: adds voice control to AI coding agents like Claude Code and Codex, letting you dictate commands and hear responses hands-free while codingLINK

tiun: unifies auth, payments, customer data, and analytics into one installable backend, letting AI and SaaS teams launch paid products quickly.LINK

Humalike x Hermes: gives AI agents turn-taking, timing, and memory APIs to decide when to speak, wait, or interrupt, useful for game characters and roleplay.LINK

Kilo Code for JetBrains: an open-source AI coding agent for JetBrains IDEs that lets you build, ship, and iterate with flexible model choices.LINK

Kilo Code for iOS and Android: mobile companion for the open-source AI coding agent, sending push alerts when sessions need input and letting you dictate prompts remotely.LINK

ditto: converts any public URL into clean Next.js or Vite code, deterministically extracting components, design tokens, fonts, and hover states via a free API.LINK

📚 Trending papers & reports

Coding AI training pairs a code writer with a trained reviewer that predicts and explains test failures, boosting accuracy across eight benchmarks while using fewer training rounds and costly test runs.LINK

Underwater robot vision gets an automated stress-test that tweaks images just enough to trip up AI scene-recognition systems, exposing where they fail before they are trusted to steer autonomous submarines.LINK

Open-source code tracing maps how reused C/C++ code spreads through chains of projects, catching ~85% of reuse paths versus ~23% for prior tools, and surfaced 48 unpatched security flaws, 23 now fixed.LINK

Digital asset profiles can now be automatically scored for whether a factory's standardized machine records are complete and usable for a given software service, flagging exactly which data is missing or mismatched.LINK

Digital twin scoring introduces a maturity model that lets factories rate and compare their digital asset records on a common scale, making it clear which ones need more work.LINK

See you tomorrow for a new dose of ☕️ Devshot!