figuring out what people actually want to use - ai, product, design. IIT Bombay. Mumbai.
Mumbai
- Tweets 254
- Following 462
- Followers 70
- Likes 214
meta just launched muse wonder 🚀
a new frontier model built for personal and professional agents, with active memory and active focus
it's live today through the meta model api and muse code
🤔 what would you build first with an agent that remembers you?
14
google just open-sourced embeddinggemma 2 🔥
one 740m model puts text, code, images, video and audio in the same search space
reportedly it runs fully on-device, apache 2.0
🤔 what would you build with search over your own voice memos?
22
anthropic just put claude managed agents in public beta 🚀
one lead agent can reportedly fan out to up to 1,000 sub-agents in a single run
that's a whole team working in parallel from one prompt 🤯
what would you throw 1,000 agents at?
16
vercel just launched eve 🚀
an ai agent is now just a folder: markdown for instructions, typescript for tools
reportedly you can deploy your first one in under 60 seconds
🤔 what would you build first with it?
1
18
amazon's alodlm gives hard tokens more compute and lets easy ones exit early ⚡
it's a diffusion language model, 1.7b and 8b, cc by-nc 4.0
reportedly 2.7x faster than vllm-served qwen3-8b on one b200
🤔 could per-token adaptive compute make diffusion lms practical?
22
sperid labs open-sourced iris-3b 🎨
a 3b image diffusion model that skips the vae and works straight on pixels, apache 2.0
reportedly it matches qwen-image on oneig, but 4x restoration doesn't beat the latent baseline
🤔 is vae-free generation the future or just a research detour?
11
unitree open-sourced a 6b humanoid model 🤖
unifolm-wla-1.0 covers 64 tasks, apache 2.0, fine-tunes on one 24gb gpu
but no per-task success rates are published
🤔 how do you trust a robot model with no track record?
12
hit researchers open-sourced sparseengine 🧩
an inference engine built for sparse attention, so long agent histories don't choke kv cache
reportedly 2.5x faster decoding than vllm at matched concurrency
🤔 will serving engines go sparse-first as agent contexts grow?
11
aleph alpha open-sourced kolibri-1 🇩🇪
78b total params, only 3.46b active per token, apache 2.0
but only the weights are open - training code and data recipes aren't
🤔 is "open weights" enough if you can't see how it was built?
15
cursor 3.0 moves from editor to an agents window 🧠
its /best-of-n command runs one task across several models in separate git worktrees, then you compare results
1️⃣ run in parallel 2️⃣ compare diffs
🤔 would you trust best-of-n over picking one model?
11
nvidia's open agent safety platform puts the guardrails outside the agent 🛡️
openshell is open source, sentry runs on a bluefield-4 dpu, per one report
idea: an agent can't talk its way past a check it can't see
🤔 is infra-level enforcement the right place for agent safety?
14
anthropic's new oss scanner sends free security reports to open-source maintainers 🔍
no human reviews them first. in its early test, 85 of 97 critical/high findings held up
so maintainers do the triage
🤔 would you opt in to unreviewed reports?
18
gemini api reportedly re-routed 3.7 flash calls to 3.8 flash without a code change ⚠️
same price per the release notes: $0.75 in, $3.75 out per 1m tokens
the old deep research agent shuts down oct 23
🤔 do you pin model versions in prod?
1
18
claude managed agents now run up to 1,000 sub-agents in one job 🤖
anthropic's own test: 66 of 70 planted bugs found, vs 14-27 for one agent
single benchmark, run by the vendor
🤔 would you pay the extra tokens for that?
1
15
reflection's first model, beam, is built for coding 🧑💻
company numbers: 65.5 on swe bench pro vs 62.1 for glm-5.2
weights are promised for later this month, not out yet
🤔 would you pick an open model you can't download yet?
1
14
qwen-image-2.1-turbo reportedly shipped this week 🤔
same 7b open-weight image model, 8 denoising steps instead of 40
that is about 5x fewer passes per picture on the same hardware
would you trade a few steps for speed on your own gpu?
1
9
tencent cloud open-sourced teamai today 🤔
team skills, rules and mcp configs live in a git repo, reviewed like code, then synced to each member's agent
it supports 16 agents, including claude code, codex and cursor
would your team review agent skills like pull requests?
2
1
17
prime intellect says 2,000 agents rewrote its coding agent in rust 🤔
over two weeks, 10,000+ sandboxes, 200 billion+ tokens (company figures)
separate agents did the review and verification, not the one that wrote the code
would you trust a rewrite checked by different agents?
8
synthesia launched syren in beta 🤔
one prompt, a document or a youtube link in, a finished video out: motion graphics, avatars, voiceover, music
videos up to three minutes, rendered in the browser
would you start a video from a document instead of a script?
17