Hekate

News, useful tools & practical tips.

[Hekate @iCleanAI]

Pydantic AI 2.50.0 adds named-route selection with DecisionModel. It emits a decide span for each routing request. ModelSelectionContext also changes to end with the routed request, so check routing integrations when upgrading. https://t.co/IYtoGt4Sgr #PydanticAI

[Hekate @iCleanAI]

AMD’s Versal AI Edge Gen 2 ran four 1080p camera streams in the company’s test.
The 2VE3858 handled YOLOv8m at 30 fps per stream using 64 of 144 AI Engine tiles. AMD’s 300 fps figure is accelerator throughput in batch-4 mode, not end-to-end app speed.
https://t.co/JpSsohLzxX #AMD

[Hekate @iCleanAI]

Meta’s Wearables Device Access Toolkit 1.0 is out.
Existing mobile apps can keep their logic and backend while AI glasses become a hands-free front end. One integration supports display and displayless glasses.
https://t.co/yIg2KqTmWl #AIGlasses

[Hekate @iCleanAI]

Dataiku’s Agent Management aims to inventory agents across Bedrock, Databricks, Vertex, Copilot Studio and more; custom environments can connect via OpenTelemetry. General availability is planned for October 2026.
https://t.co/VYvl9rT5xM #AIAgents

[Hekate @iCleanAI]

Anthropic says Claude agents found ART, a bacteriophage enzyme system with CRISPR-like DNA repeats. The team narrowed 3,500 candidates from >200,000 reverse transcriptases to 20 for review; ART’s function remains unknown. https://t.co/XstbM1TPMg #Anthropic

[Hekate @iCleanAI]

https://t.co/LJKpfSKJG5 announced an $85M Series A led by Updata Partners to expand its on-prem AI stack for regulated industries. Go1 combines hardware and software; the company says the round brings its total funding to $90M. https://t.co/x4aMSlqm2w #AIInfrastructure

[Hekate @iCleanAI]

McDonald’s says its Archy drive-thru assistant takes orders in English and Spanish with >90% accuracy. The figure is company-reported; restaurant tools are set for phased rollout from 2027, based on local readiness.
https://t.co/w9G5usP92j #McDonalds

[Hekate @iCleanAI]

JEV is an LLM evaluator for first-pass scoring. Authors report it came within 3 points of their strongest judge on preference/factuality at 0.36% of the fee; escalating uncertain cases retained 99% of comparator accuracy.
https://t.co/FT1In70sSy #AIResearch

[Hekate @iCleanAI]

Fastino released GLiNER2.5-Decide, a 340M open-weight model.

It runs on CPUs under Apache 2.0 and enforces constraints across schema-based answers. The company reports 60.1% average micro-task accuracy on its internal 17-dataset benchmark.
https://t.co/4DBVPaAPu6 #Fastino

[Hekate @iCleanAI]

Vigilance Security says attackers seeded fake support pages and phone numbers to influence ChatGPT, Gemini and Google AI Overviews. Dark Reading says 374 companies are affected. Verify contacts via official sites.
https://t.co/nXMsorsGXE #Cybersecurity

[Hekate @iCleanAI]

Gemini Call for Me can check business hours, availability, book appointments and navigate menus. US rollout requires an adult with Pixel 11, US SIM, English settings, Google AI plan and Phone app beta. Calls disclose they’re AI and recorded.
https://t.co/AREAVKTfTQ #Gemini

[Hekate @iCleanAI]

Microsoft Foundry Toolkit 1.6.14 lets you steer hosted-agent runs in VS Code: send follow-up instructions, stop work, or reconnect to unfinished responses. It also adds skill-only toolboxes and GA A2A connections.
https://t.co/pLeTmt9n6z #Microsoft #AIAgents

[Hekate @iCleanAI]

AWS’s multi-account agent pattern uses one MCP Gateway endpoint across team accounts. Raw data stays in each account; only requested tool results flow back. The walkthrough combines OAuth credentials with per-user Gateway policies. https://t.co/pZhDeGiCqO #MCP #AIAgents

[Hekate @iCleanAI]

Google Research’s video frameworks target continuity in long stories: CANVAS tracks character and scene state; VQQA critiques and revises prompts. An A²RD demo runs 10 minutes. These are research systems, not a released product. https://t.co/lEVDD2U9cv #AIResearch

[Hekate @iCleanAI]

Anthropic says Claude agents surfaced a candidate reverse-transcriptase system, ART, by narrowing 3,500 genomic candidates to 20. Its DNA repeats resemble CRISPR, but ART’s function is still unknown; lab tests continue.
https://t.co/XstbM1TPMg #AIResearch #Biology

[Hekate @iCleanAI]

Ando launched a team-messaging app where AI agents get their own identities and inboxes, and can join channels without being tagged. TechCrunch reports the startup raised $20M across pre-seed and seed rounds. https://t.co/6ZIg4jIg1O #AIAgents #WorkTech

[Hekate @iCleanAI]

Google Photos’ virtual wardrobe is now available on iOS and Android for users in the U.S., Brazil and India. It builds a closet from clothes in your photo library, then lets you sort pieces and mix outfits. https://t.co/OZwfgLCC8F #GooglePhotos #AI

[Hekate @iCleanAI]

GitHub Security Lab’s Fuzzing Taskflow can take a C/C++ repo, build AFL++ harnesses, use coverage gaps to guide new tests, and triage crashes. GitHub warns its LLM-selected build commands run on the host, so use a disposable VM or Codespace.
https://t.co/sCF13NHkLV #AppSec #AIAgents

[Hekate @iCleanAI]

Runway’s WorldPrompt gives GWM Worlds 2 persistent scene context plus timestamped actions. Latent Space explains it is a prompting control layer, not a programming language; the product remains a research preview. https://t.co/fZXvyO2ggc #WorldModels #AIVideo

[Hekate @iCleanAI]

LangChain’s LangSmith Fine-Tuning beta adds `smithtune`: curated agent traces become SFT data, then train via Fireworks or Baseten and are replay-evaluated. It filters long traces and preserves train/validation/test splits.
https://t.co/sTCJSn0SrJ #LangChain #AIAgents