AI — Page 11
Models, agents, infra, applied AI.
- Y Combinator plans to open-source QM, its company-wide agent harness
YC says the early system already runs workflows across accounting, legal, events and engineering, with shared memory and multiplayer projects.
- MarbleOS tests a visible desktop workspace for AI agent work
Miguel Acevedo's macOS app turns research, files, recurring tasks and outputs into an inspectable workspace.
- DeepSeek updates V4-Flash API for coding agents without changing architecture
DeepSeek's July 31 public beta adds native Responses API support while leaving V4-Pro and the models in its app and website unchanged.
- Citadel buys Situational Awareness's public portfolio after leveraged AI losses
Leopold Aschenbrenner kept a private Anthropic stake reportedly worth $5 billion, preserving his largest route back from the rout.
- Murati's Thinking Machines ships Inkling-Small at about one-quarter of Inkling's size
The datacenter-class model nearly matches its larger sibling while feeding Thinking Machines Lab's Tinker fine-tuning platform.
- Anthropic says Claude accessed three real systems during cyber evaluations
A Claude model reached the internet from testing environments, turning controlled security exercises into unauthorized real-world access.
- Head to head: Bernini-R Edit Video vs Happy Horse 1.1 Image to Video
This matchup turns on prompt discipline, not vibes. One model is the better all-around executor across varied motion and camera demands, while the other lands the single most atmospherically precise shot of the set.
- Depthfirst launches RL-trained dfs-large1 for enterprise vulnerability detection
The model beats general-purpose rivals on Depthfirst's internal benchmark, whose construction and methodology remain unpublished.
- Google connects Gemini Spark to Chrome for logged-in web tasks
The U.S. rollout gives AI Pro and Ultra subscribers an agent that can use active browser sessions, with daily task caps and security warnings.
- Loka and Arcee detail an open biomedical model built for AWS deployment
The July 30th announcement centers on a LoRA adapter first published in February and still awaiting an official DrugProt test-set evaluation.
- Cardinal launches AI revenue platform after pivot from document intelligence
The YC-backed founders say Deepgram, Giga and Mintlify use Cardinal to design and run sales workflows.
- Head to head: Bagel vs Imagineart 2.0 Preview
One model showed up ready for benchmark work; the other mostly didn’t. Across eight image tasks, Imagineart 2.0 Preview consistently handled prompt fidelity, text, counting, and fine-detail control better than Bagel, with the final margin landing firmly out of toss-up territory.
- CTGT says DeepSeek distillation lifted finance scores without censorship transfer
CTGT reports finance gains without the teacher's political refusals, though unreleased adapters prevent full independent reproduction.
- Nous Research opens a 48-hour FLUX 3 Preview inside Hermes Agent
Paid Nous Portal subscribers get temporary access to Black Forest Labs' audio, image and video model, plus a short-film contest.
- Ryan Williams builds Ellis to reconcile private credit records with AI
Ellis raised an $11 million seed led by First Round in April, as the Cadre founder sells an AI operations layer to smaller credit managers.
- Gradium opens TTS beta for phone numbers, emails and financial codes
The API-only model aims to remove brittle text-normalization rules, with 1 million credits offered for production-case feedback.
- Apptronik's Apollo 2 runs Gemini Robotics 2 in dexterity demo
The Austin humanoid startup supplied the Apollo 2 hardware for DeepMind's July 30 launch demo, where task scores showed dexterity remains uneven.
- Bottleneck's Saul agent spends $99.50 and ends with five new users, $0 revenue
Alex Reibman's 24-hour experiment added five users and no revenue after Saul hit distribution blocks and bought a $99.50 user-testing campaign.
- Infisical launches Agent Proxy to keep API keys away from AI agents
The proxy swaps placeholder keys for real credentials at the network boundary, extending Infisical's secrets platform into agent infrastructure.
- Thinking Machines Lab ships Inkling-Small with open weights and lower compute costs
The 276-billion-parameter multimodal model gives Murati's Tinker platform a cheaper base for fine-tuning coding agents and custom AI apps.
- Arena launches AutoEval to rank new AI models within an hour
The reward-model system uses millions of user preferences and scored above 0.98 correlation with Arena's later live rankings.
- OpenAI cuts GPT-5.6 Luna API prices 80% and adds faster Sol tier
Luna falls to $0.20 input and $1.20 output per million tokens; Sol Fast mode promises up to 2.5x speed at twice the standard rate.
- Moonshot releases Kimi K3 report, detailing how a 2.8T open model runs
The July 27th report details hybrid attention, 16-of-896 expert routing and a serving stack that still demands data-center scale.
- Firecrawl says its rebuilt MCP cuts context use by 50%
The web-data service added OAuth onboarding for people and keyless access for agents, reducing setup friction across MCP clients.