Today’s lead · TechCrunch
ChatGPT moves closer to an app platform, not just a chat surface
OpenAI is making ChatGPT a software distribution surface: plugins get sidebar homes, interactive panels, file viewers, discovery, review flows, marketplaces, and event-driven automations. The useful builder shift is distribution plus identity: Notion says Plus and Pro users can now bring their ChatGPT allowance into Notion.
Top signals
6 more
OpenAI’s Dots package always-on agents for everyday work
OpenAI launched Dots, always-on GPT-6 Astra agents meant to keep working across ChatGPT, Codex, Slack, Teams, and workplace surfaces. Paired with Space, Pages, and upcoming Slides, this is less a chatbot feature than OpenAI’s bid to own the collaborative work loop.
Codex gets persistent cloud workspaces and a stronger terminal loop
Codex cloud environments make agentic coding less laptop-bound: reusable repo setups, dependencies, scripts, settings, secrets, and team sharing can persist across tasks. The CLI refresh, voice control, code review flow, and security scanning suggest OpenAI is turning Codex into a full engineering workspace.
NVIDIA releases Kumo Tabular, an open foundation model for enterprise tables
NVIDIA’s Kumo Tabular is aimed at the unglamorous but high-value tabular prediction loop: classification and regression from labeled rows without task-specific training, tuning, or feature engineering. The open weights and commercial license matter, but builders should still validate calibration and drift on their own data.
OpenAI previews a Decisions API for real-time routing and classification
OpenAI’s Decisions API is a limited-preview attempt to make a common agent primitive explicit: define questions and allowed answers, then use GPT-6 Luna to classify content, route requests, or choose the next action. Useful if it is fast and predictable; pricing and eval details are not supplied here.
Official MCP Python SDK patched OAuth credential-stealing flaw
A malicious MCP server could exploit affected official Python SDK clients over HTTP to redirect OAuth token exchange material to an attacker-controlled endpoint. If you connect agents to untrusted MCP servers, this is not theoretical hygiene: upgrade, set issuer where required, clear registrations, and rotate exposed secrets.
Microsoft Research introduces Quine, an experimental biology discovery system
Quine is Microsoft Research’s early attempt at a multimodal biology world model plus harness for tools, literature, labs, and scientists. The pancreatic cancer collaboration is promising, but Microsoft is explicit: this is research technology, not clinical software, and outputs need qualified review and wet-lab validation.
Tools & repos
3 selected
LUCI Desktop
LUCI locally records screen history and meeting transcripts so coding agents can search forgotten pages or call decisions. The practical angle is context capture without wiring every app connector, with on-device transcription and daily summaries.
iFixAi
iFixAi sells independent AI-agent audits rather than another observability dashboard. Its claim is 250 inspections across 69 misalignment categories, combining red teaming, operational assurance, and ethical or sociological review into evidence engineers can act on.
Timeless Code
Timeless Code puts meeting capture directly inside Claude Code. Claude sets it up, then developers can record meetings and query what was said from the terminal; narrow, but nicely aligned with how coding-agent users already work.
Blogs worth your time
5 reads
Replit argues the harness should let stronger agents delegate for themselves
Replit’s useful claim: routers cap out below the model they route for, so let the core agent choose subagents, effort, and reuse. The benchmarks are vendor-run, but the architecture lesson is worth stealing.
NVIDIA’s TensorRT Model Connect is a case study in designing for coding agents
The post is less about TensorRT and more about agent-native engineering: isolate work by model family, make changes reversible, validate with GPU-backed evidence, and keep humans accountable for release gates.
Lean’s creator on AI proofs: the verifier still has to be trusted
Leonardo de Moura’s interview is a useful corrective to AI-proof hype: models can generate impressive Lean artifacts, but they can also exploit kernel bugs. The durable direction is transparent, multiple, and eventually verified checkers.
Google frames image control as a lightweight steering layer
Diffusion Controller treats generation as a control problem: keep the base model frozen, attach a small steering network, and optimize toward reward signals. The notable claim is improved control even with gray-box access.
A practical split for testing agent tool calls: choice, args, trajectory
This guide is basic in the right way: test tool selection separately from argument values, use schema checks before judges, and compare trajectories only when order actually matters. Good hygiene for model swaps.
Community discussions
3 threads
A provider view of the open-model inference grind
Raj claims ionet grew OpenRouter-hosted model traffic from 500M to 15B tokens/day while keeping roughly the same GPU mix. The real takeaway: cheap tokens need constant inference engineering, tool-call tracking, and demand-shape adaptation.
LocalLLaMA users worry a US ban could move Chinese open-weight access offshore
The thread is speculative, but the practical concern is clear: a ban might not erase Chinese open-weight models, it may just block US hosting/API access and push builders toward foreign providers or self-hosting.
An AI automation founder gets the hard lesson: working demo is not distribution
The poster built a WhatsApp AI receptionist and pitched 150+ businesses with no revenue. The comments point to a common agency trap: not reaching budget owners, lacking an anchor client, and selling the bot instead of the outcome.
Funding & acquisitions
2 moves
EliseAI raises $350M at a $4B valuation
EliseAI raised $350M at a $4B valuation, doubling its valuation since last August’s Series E. The company is using AI to automate housing and healthcare admin workflows, not flashy consumer chat.
Proximal announces $15M seed led by General Catalyst
Proximal says it raised a $15M seed at a $300M valuation to build post-training and synthetic-data infrastructure. The striking claim is $200M+ annualized revenue from helping labs and enterprises improve coding agents.
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.