And here you go ! After weeks of work, my open access course on machine unlearning is live. ๐
Open-access course link here - https://t.co/AVC0pCG6Hd .
Also check out the GitHub repo on MUL with all the colab notebooks - https://t.co/2DPoWY5dip
Bring Responsible AI to everyone !
#ResponsibleAI #MachineUnlearning #OpenAccess #github #rai #ai
@_akhaliq speculative decoding for VLMs makes sense, image tokens eat so much when agents read screenshots all day. gonna give this a detailed read this week
@neural_avb running it free in the browser is such a good idea tbh. people with scanned bank statements or invoices don't want to upload them anywhere ! would love to use this too ๐ฅ
Fable 5.2 wrote the transformer.
Opus 5.5 generated the training data.
The result:
a 367K parameter language model that fits on a floppy disk and runs on a Pentium.
AI is now building AI for hardware from 1995.
Been talking to a lot of Indian founders building AI products lately, through @BuildersQuorum and the Claude dev community.
One pattern keeps coming up. The demo works great, then real users show up and do stuff nobody planned for ๐ญ
Founders, what's the one thing real users did that completely broke your assumptions? Drop it below, wanna learn from you guys !
Yepp and the tricky part is the run still "works", loss goes down so nobody questions it. Like on a small image dataset it'll go straight for a big pretrained backbone + heavy aug when a small conv net would've done the job in minutes. Making it write a dumb baseline first catches most of this
This is honestly sooo good from @GnaniAi ๐ฅ
Anyone who's built voice stuff for India knows the pain. People switch between Hindi and English mid sentence, accents change every few hundred kms, and half the calls are super noisy. Most speech to text models just break there ๐ญ
Prisma v3.0 is built exactly for that, and 9 of its 10 languages are Indian !
Builders, let me know in the comments, how bad has mixed language been for your transcripts ?
https://t.co/3dDScSkKl0
India rarely speaks one language at a time.
Meet Prisma v3.0, our new speech-to-text (STT) model, built for the way India actually talks. It handles Hindi and English in the same sentence, regional accents and noisy 8 kHz phone calls.
It covers 10 languages, including 9 Indian languages, with four output modes and keyterm prompting so brand names, product names and numbers come out right. That means transcripts your team can use as they are.
Try it out. https://t.co/rK2jAwEHmE
Damnn the shadows off the boxes stay properly consistent for 9M params ๐ฅ
Went through the repo, train_wm.sh has K=4, so at 30fps the model only ever sees the last ~0.13 sec. Whatever long term continuity it has must be coming from what's still visible in those 4 frames, and it's a dark room, so the moment the flashlight turns away a box is basically gone from the context.
Test I'd add to the 60 step dreams in select_ckpt - park the beam on a box, turn away for 2-3 sec, turn back. Same spot and angle means the VQ tokens are carrying way more scene info than 4 frames should allow. If it shifts or vanishes thats where the drift is
Keeping them running in a Space after I close the laptop is the bit that actually matters for the 24/7 thing deepfates is asking about.
Although from the README, sessions can share a folder and groups don't isolate files or creds, so it's sandboxed from my machine, not from each other. Real case - Claude Code is halfway through a refactor, Codex is reviewing in the same checkout, Space restarts at 3am. Both get told their tool calls were interrupted and pick up by themselves, but on top of whatever half written state the other one left behind.
Does each session get its own git worktree or is it one checkout for everyone ?