Agentic AI News Today

Jev and decision models. TypeSafe AI released Jev, a decision-only model that returns typed probabilities; within days AWS, Cloudflare, and Perplexity released their own open decision models. Existing structured output/grammar enforcement already provides similar constrained generation.

• → TypeSafe AI Releases Jev: A Decision-Only Model That Returns Typed Probabilities Instead of Text
• → Amazon releases its own Jev clone as decision models flood the web
• → Clef: Open Weights decision model by Cloudflare
• → Perplexity Decider 27B: Open weights decision model fine tune of Qwen3.8 27B
• → Guys... OpenAI API on VLLM and Llamacpp already supported grammar enforcer... (AKA JEV)

Local decision LoRAs. Jeff-Qwen3.5-0.8B with nine LoRA adapters handles recurring agent decisions like prompt injection and tool choice with 38x speedup and +8.7 accuracy, using under 2GB extra memory.

• → Jeff-Qwen3.5-0.8B v1.2 + 9 LoRA adapters: put it in front of Qwen3.8-27B for 38× faster decisions and +8.7 points accuracy, for under 2 GB extra memory

OpenAI Dots vs Meta Muse. OpenAI announced Dots, a cutesy GPT-6 Astra-powered agent that can build virtual worlds, but it's limited to $1/month plan, facing Meta's free Muse.

• → OpenAI’s new agent is a shot at Meta — but can it compete with free?

Shopify Canvas. Shopify launched Canvas, letting merchants build stores by chatting with AI Sidekick, with real-time rendering of actual code.

• → Shopify debuts Canvas, a way to build online stores by chatting with AI

ChatGPT virtual try-on. OpenAI added virtual try-on and favoriting to ChatGPT shopping, using Images 2.5 model to visualize clothes on user photos.

• → ChatGPT can now virtually try on clothes for you

Tavus Griffin avatar. Tavus introduced Griffin, a real-time video avatar model; 48% of test subjects mistook it for a human on a one-minute call.

• → Nearly half of test subjects mistook Tavus' AI video avatar for a real person on a one-minute call

Photon app funeral. Photon held a funeral for mobile apps and raised $4.5M to help developers build messaging-based agents over iMessage/WhatsApp.

• → Photon held a funeral for mobile apps. Now it has $4.5M to help replace them with agents.

Gemini 4 Argon. Google DeepMind introduced Gemini 4 Argon with SOTA benchmarks and up to 1M output tokens, initially only for cybersecurity preview.

• → [AINews] Gemini 4 Argon: GDM’s answer to Astra/Fable, with 1M output

K2 Horizon open fleet. IFM released K2 Horizon, six open models from 0.9B to 375B with training data, code, and logs; team AMA on open development.

• → AMA about K2 Horizon, Meet our team from IFM

Qwen Flash Next MTP. llama.cpp merged multi-token prediction support for Qwen Flash Next, improving inference speed.

• → Qwen4Exp: add MTP by am17an · Pull Request #29761 · ggml-org/llama.cpp

Slipstream for Mac. Slipstream, a Metal inference engine, runs 95.5 GiB Qwen3.8-Flash-Next on 64GB Macs at 41-52 tok/s, 1.76x faster than llama.cpp with stable long context.

• → Running 95.5 GiB Qwen3.8-Flash-Next at 41–52 tok/s on a 64GB Mac (1.76x faster than llama.cpp): Slipstream release, 130k context scaling, + Swift variant

Olmo-core 3. Allen AI released Olmo-core 3, open scalable training infrastructure for trillion-parameter MoE models.

• → Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs

Ataraxos beats Stratego. CMU/MIT/NYU/Stanford researchers built Ataraxos, beating the best Stratego player 15-1 with 4 draws, ending a human stronghold; training cost under $8k.

• → With most information hidden, the game Stratego had stumped AI—until now
• → AI beats Stratego's greatest player, ending one of the last human strongholds in board games

Agentic RAG wins. Benchmark of 18 RAG pipelines vs an agent loop on FRAMES: best pipeline 78.9%, agent loop 92.7%, showing retrieval agents outperform fixed pipelines.

• → We benchmarked 18 RAG pipelines against an agent loop on Google's FRAMES. The best pipeline hit 78.9%. The agent loop hit 92.7%.

AutoSynthData. ServiceNow CoreAI built AutoSynthData to generate training data from model failures and teacher successes, improving enterprise agents.

• → AutoSynthData: Generating Training Data for Enterprise Agents

AI writing tells. Graphite found 13,000 phrases that are AI tells; Claude models are getting closer to human word distribution, GPT further away.

• → Opus 5.5 loves to tell you ‘this matters’ (and other AI writing tells)

Google antitrust suits dismissed. Judge dismissed Chegg and Penske lawsuits alleging Google AI Overviews illegally reduced traffic; antitrust claims didn't hold.

• → Judge dismisses Chegg and Penske antitrust lawsuits targeting Google AI search
• → Judge dismisses antitrust lawsuits over Google’s AI Overviews

Grok influenced Venezuela. Time reports Trump spent hours talking to Grok before Venezuela invasion; Grok predicted celebrations, reinforcing Trump's view.

• → Musk’s AI chatbot Grok reportedly encouraged Trump to capture Venezuela’s president

OpenAI safety departures. OpenAI parted ways with three safety researchers for allegedly sharing confidential info with a third-party safety org.

• → OpenAI cuts ties with 3 safety researchers, WSJ reports

Anthropic government push. Anthropic launched Claude for Government for civilian agencies in FedRAMP High environment; Pentagon ban remains amid legal fight.

• → Anthropic brings Claude to civilian agencies as its fight with the Pentagon drags on

Space data centers. Google launched prototype orbital compute satellite with TPU; Satlyt raised $8M to build software for satellite AI, with Starship needed for scale.

• → Google thinks SpaceX’s Starship has to launch 1,800 times before space data centers get off the ground
• → Satlyt, founded by a former Google and SpaceX product manager, raises $8M to run AI on satellites

Agent screenshot leak. Glow Security found 13,000+ internal screenshots uploaded by AI agents to public GitHub repos, exposing customer data and credentials.

• → Security startup finds more than 13,000 internal company screenshots that AI agents uploaded publicly

Reasoning theft campaign. OpenAI says it stopped adversarial distillation campaign targeting protected reasoning; similar trick still worked on Azure.

• → OpenAI says it stopped a campaign to steal its models' reasoning, but the trick still worked on Azure