Sign in

Kabir Kumar, aiplans.org

@kabirkumar.bsky.social
1.3K followers 1.5K following 5.2K posts

I run AI Plans, an AI Safety lab focused on solving AI Alignment before 2029. For several weeks I used a stone for a pillow. I once spent a quarter of my paycheck on cheese. Ping me! DMs not working atm due to totalitarian UK law :( SurpassAI

PostsRepliesMedia
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 15/09/2026
rightly so
010
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 11/08/2026
Just fired someone for the first time today. It feels absolutely horrible. For Donald Trump to have done it so much and so easily, that it became his catchphrase is insane. How does a person do that while having empathy in their heart?
010
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 10/08/2026
evidence?
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 10/08/2026
240
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 10/08/2026
sent this to my gf
020
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 09/08/2026
crazy to think that there's people alive today who were still around when cameras just took black and white photos
010
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 08/08/2026
false
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 08/08/2026
wtf
010
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 08/08/2026
an ai company wouldnt take this much responsibility
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 08/08/2026
sorry, i changed likes no. from 67
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 19/07/2026
sorry, a bit too news spammy just got annoyed by this article just literally referencing the author's own twitter as a source
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 19/07/2026
The Guardian is such a joke i see this headline: but the source is on twitter and it's literally the author referencing himself
100
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 18/07/2026
Fable hallucinates and continues doubling down even after 3 messages to the contrary. It hallucinated the line "check gpt 5.6's suggestions for a 5 day version and consider whether the two of you disagree anywhere" and kept doubling down even when contradicting evidence was shown.
010
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 08/07/2026
100
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 08/07/2026
Screenshot from Science journal site. paragraph says:

Security and Safety Concerns
Authors and reviewers are expected to notify editors if a manuscript could be considered to report Dual Use Research of Concern (DURC). Gain-of function experiments are considered under our DURC policies. The National Science Advisory Board for Biosecurity has defined DURC as “life sciences research that, based on current understanding, can be reasonably anticipated to provide knowledge, information, products, or technologies that could be directly misapplied to pose a significant threat with broad potential consequences to public health and safety, agricultural crops and other plants, animals, the environment, materiel, or national security.” Papers identified as possible DURC risks, or analogous risks in the physical sciences, will be brought to the attention of the Editor-in-Chief for further evaluation. If necessary, outside reviewers with expertise in the area will be consulted.
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 08/07/2026
www.whitehouse.gov/presidential...
whitehouse.gov
Improving the Safety and Security of Biological Research
By the authority vested in me as President by the Constitution and the laws of the United States of America, it is hereby ordered: Section 1.  Purpose.
100
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 08/07/2026
bidenwhitehouse.archives.gov/wp-content/u...
bidenwhitehouse.archives.gov
100
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 08/07/2026
also, 'forbidding the open circulation of language' is how it's *already* done - e.g. www.congress.gov/crs_external...
congress.gov
100
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 08/07/2026
one way to do this is to ban training on gain of function research/bio/cyber capabilities - e.g. this was done for GLM 5.2
100
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/07/2026
hmm, maybe authors get like, a limited number of initial free submissions?? like, first 5 submissions are free and guaranteed some amount of peer review? and if nothing valuable is found in the paper, then next time, they have to pay for review, with rising costs each time its low value?
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/07/2026
just ones that have stuff like made up citations, claim to be inventing something new that's obviously not, bunch of obvious errors, etc - just the big obvious stuff
110
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/07/2026
yeah, afaik, its hard to actually tell when that is, so when it isn't, its just adding friction to work that isnt that, which is the majority - so, making another market for labs without that friction. my knowledge of this just comes from skimming some articles and talking with a biosecurity friend
110
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/07/2026
funny if true
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/07/2026
i fucking hate this company
040
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/07/2026
use that to pay peer reviewers
010
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/07/2026
Have a fine for rejections that were due to AI/slop type stuff
210
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/07/2026
yes, smartest guy i knows says stuff in super simple ways to me
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/07/2026
at double the token limit price
020
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/07/2026
this sounds like nothing and is embarassingly slow. better than literally nothing ig
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/07/2026
seems false and to be about manus acq stuff
020
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 29/06/2026
SQUACK\!!!
010
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 29/06/2026
france was like a USA + China combowombo
Heads of aristocrats on pikes.
Robespierre. Montesquieu's The Spirit of Law defines a core principle of a democratic government: virtue—described as "the love of laws and of our country."[11] In Robespierre's speech to the National Convention on 5 February 1794, he regards virtue as being the "fundamental principle of popular or democratic government."[12][13] This was, in fact, the same virtue defined by Montesquieu almost 50 years prior. Robespierre believed the virtue needed for any democratic government was extremely lacking in the French people. As a result, he decided to weed out those he believed could never possess this virtue. The result was a
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 29/06/2026
fucking banger
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 29/06/2026
I think mine is "avoiding messaging people who are happy to help and who would be happy to receive my message because i'm scared of messing things up"
020
Reposted by Kabir Kumar, aiplans.org
norvid_studies @norvid-studies.bsky.social · 28/06/2026
from @cee.wtf
219926
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 29/06/2026
what about e.g. 5 mentally unstable working together?
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 29/06/2026
also, there's more and more automated bio labs, a bunch of whom don't really seem to do thorough background checks, or thorough checks in general. and they're growing
100
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 29/06/2026
yup
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 29/06/2026
all gamba
010
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 24/06/2026
which one was the best?
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 24/06/2026
Yeah, tried GLM 5.2 on a task that both Opus 4.8 and GPT 5.5 were being really fucking annoying at. And Gemini 3.1 pro, as well. It did much much better. Actually did what I wanted it to do. It's a good model.
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/06/2026
unsure if openai is worse, now. i think anthropic is just much betrer at marketing and lying
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/06/2026
you're very welcome! Would be excited to hear what you think if you read it!
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/06/2026
You Are Special by Fred Rogers truly, truly beautiful, one of my all time favourite books by a beautiful person
110
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/06/2026
hmm, Concepts of Modern Mathematics by Ian Stewart
110
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/06/2026
oh whoops, *other* than ai
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/06/2026
three liability regimes for artificial intelligence by beckers & teubner
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/06/2026
hes also very anti ai datacenters
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/06/2026
It took me 4 tries to guess who @krissy.bsky.social was @mystery-mutual.bsky.social mystery-mutual.fly.dev
mystery-mutual.fly.dev
Mystery Mutual
Guess which of your BlueSky mutuals these statements describe.
000
Kabir Kumar, aiplans.org @kabirkumar.bsky.social · 07/06/2026
what was the statement?
100