Today’s lead · TechCrunch
Trajectory opens its post-training platform for domain experts
Trajectory is pitching continual post-training as something lawyers, doctors, and analysts can steer without an internal research team. The useful bit is the feedback loop around real applications; the risk is treating “keeps learning” as automatic magic before teams understand evaluation, governance, and failure modes.
Top signals
7 more
Meituan introduces LongCat-2.5 for multimodal agent work
LongCat-2.5 is Meituan’s latest natively multimodal agentic model, with a claimed 1.6T total parameters and about 48B active. The pitch is long-horizon work across coding, visual coding, creative design, and 3D tasks; the dossier provides launch claims, not independent benchmark detail.
Critical Langflow flaws put agent servers on the patch-now list
Langflow users should treat this as operational debt, not abstract CVE noise. IBM reported critical flaws affecting releases from 1.0.0 through 1.12.8, including unauthenticated auth bypass and arbitrary code execution paths. Version 1.13.0 is the stated fix, with no IBM workaround cited.
Claude Messages API gets server-side conversation threads
Anthropic’s public beta for message threads is a practical latency and payload improvement for chatty Claude apps. Instead of resending the full transcript each turn, developers send new messages plus a previous message ID, while the API caches conversation state server-side.
MIT’s OPAL asks whether optimizers should rewrite themselves mid-run
OPAL is a provocative research direction: put an agent inside the optimizer, let it watch training dynamics, and rewrite update code during a single run. The appeal is recipe-free adaptation for new architectures or datasets; the open question is cost, reliability, and scale beyond reported tasks.
3D World Simulator brings tactile, multi-view robot rollouts into one model
3D World Simulator is interesting because it models robot interaction in a unified 3D space, not just as 2D video prediction. The demo claims 50+ FPS on an RTX 5090 and supports novel views, RGB-D, geometry, and tactile signals, while warning it is still a research demo.
OpenAssembler takes a serious run at NVIDIA’s closed PTX-to-SASS layer
OpenAssembler matters if you care about the compiler layer beneath CUDA performance. It reverse-engineers NVIDIA’s closed ptxas path for sm_80 and reports selected kernel wins, but the repo itself narrows the claim: workload-dependent microbenchmarks, Ampere targeting, and full-repository compilation still in progress.
OpenAI’s LifeVal benchmark shifts evals toward everyday assistant work
LifeVal is OpenAI’s attempt to measure assistant usefulness on messy civilian tasks: childcare, finances, government services, and paperwork. That framing is welcome for product builders, though the supplied evidence only describes the benchmark’s intent and user-feedback basis, not methodology or scores.
Tools & repos
5 selectedvercel-labs/agent-desktop
A Rust CLI for giving AI agents computer-use capabilities. The trend spike suggests builders are still hunting for lower-friction local control surfaces for agent workflows.
FlashML-org/FreeToken
FreeToken’s promise is datacenter-scale model serving on a desktop. Treat the claim as something to benchmark locally, but the developer interest is obvious.
beacon.sh
Beacon targets a real agent pain: every coding session starting cold. It turns traces from Claude Code, Cursor, Codex, OpenCode, and other harnesses into shared memory, search, and reusable lessons.
Agent Inbox
Agent Inbox gives agents one API-backed identity across email, Slack, and Telegram. It is a glue-code reduction play for teams putting agents where users already message.
Clayzo
Clayzo lets coding agents generate web animations and motion graphics in code, review frames, then export video or run live on a site. Useful if design output already lives in your agent loop.
Blogs worth your time
5 readsOpenAI explains Pando, its Git-compatible backend for Codex-scale development
The hook is not “Git is dead”; it is that agent-generated code changes the scale assumptions around repositories. OpenAI says Pando is Git-compatible and built for the development volume Codex makes possible.
NVIDIA open-sources Attribution for distributed AI failure diagnosis
This is practical infra work: turning cascades of NCCL, scheduler, GPU, and fabric symptoms into one workload-scoped diagnosis. NVIDIA says Attribution is open source and already used across its AI infrastructure.
NVIDIA shows how Cosmos moves from world-model recipes to NIM serving
Cosmos is being framed less as a single model and more as a development stack: train, adapt, optimize, and serve world models. The post is especially useful for GPU memory planning and simulation workflow selection.
LangChain argues proactive agents reopen the context-engineering problem
Harrison Chase pushes back on the idea that agent harness problems are solved. His useful list: proactiveness, memory, async work, goal tracking, alignment, and browser use become harder when agents run continuously.
Simon Willison catches the Census Bureau’s AI survey becoming an SI survey
This is a good reminder that AI adoption charts can move because survey wording moved. Willison shows two material Census Bureau changes: “producing goods or services” became “any business function,” then AI became “Super Intelligence.”
Community discussions
1 threadsLocal LLMs for city-council accountability hit the transcript bottleneck
A user loaded 16 years of city council minutes into Qwen to investigate Flock camera approval, then ran into missing meeting detail. The thread’s practical advice is boring but solid: scrape YouTube transcripts or transcribe locally with Whisper.
Funding & acquisitions
4 moves
42AI emerges with $62.5M for behavioral world models
42AI’s claim is bold: probabilistic models of individual behavior, first used with financial brokerages. The funding is real; builders should treat the predictive-human-behavior framing as high-stakes and scrutiny-worthy.
Blacksmith raises $66M as CI becomes the AI-coding bottleneck
Y Combinator says Blacksmith raised a $66 million Series C at a $1 billion valuation. The sharper point: as agents write more code, faster validation infrastructure becomes leverage, not cleanup.
Omen AI buys Aquatic Labs to merge data-center water sensors
Omen AI is buying Aquatic Labs in a $16 million cash-and-stock deal. The combined bet is that real-time cooling-fluid monitoring can reduce downtime and water waste in AI data centers.
Clinics On Cloud raises Series A to scale AI-enabled microclinics
Clinics On Cloud raised Series A capital from Model42 to expand health ATMs and microclinics. The plan is AI capabilities, R&D, manufacturing, partnerships, and using the UAE as a global hub.
Bengaluru radar
0 eventsThere are no relevant Bengaluru events to highlight today.