Ram Bharadwaj (@arbdwj) header
Ram Bharadwaj (@arbdwj) profile picture

AI safety research @lasrlabs | previously @lossfunk

London, England
Joined December 2021
Photo by Ram Bharadwaj (@arbdwj)Photo by Ram Bharadwaj (@arbdwj)Photo by Ram Bharadwaj (@arbdwj)Photo by Ram Bharadwaj (@arbdwj)Photo by Ram Bharadwaj (@arbdwj)Photo by Ram Bharadwaj (@arbdwj)Photo by Ram Bharadwaj (@arbdwj)Photo by Ram Bharadwaj (@arbdwj)Photo by Ram Bharadwaj (@arbdwj)
Ram Bharadwaj retweeted
person who thinks AI will prove the Riemann hypothesis in 2026 but that Apple won’t deliver an LLM-powered Siri until 2027
161
436
155
5,268
374,839
Ram Bharadwaj retweeted
the world is a hackathon now
27
163
20
1,185
23,024
Ram Bharadwaj retweeted
anthropic ipo has got the prestige parasites so thirsty that they're unironically posting rejection letters just to show that they were able to get an interview
27
44
14
1,671
39,394
LLMs judge AI safety research worse when they believe a machine, not a human, is grading them. On Anthropic’s TASTE benchmark, they drift more toward ideas that are easy to execute and verify, and away from what human-experts rate as more important. 🧵
1
1
16
633
I don't see anything wrong with Rahul Gandhi sharing snippets of the physical intimacy his dad & mom shared. It's moments like these that I feel let down by India's hypocrisy. Love is a lot about intimacy... Physical & emotional. You can critique RaGa in everything, but this.
7
6
55
2,770
Pet peeves for AI papers/articles: 1. The words "key," "ensure," and "increasingly." 2. The sentence: 'Frontier AI systems are used in increasingly used for X, but with this comes risks of Y.' 3. The phrase: "access its benefits while mitigating the risks."
5
2
54
3,821
Explain why large-scale human clone production is bad without moralfagging
An artificial womb just successfully grew a baby lamb. This is disturbing on so many levels and giving me serious Matrix vibes. We’re really entering the “playing God” era… next thing you know, we’ll have large-scale human clone production.
Readers added context they thought people might want to know
Artificial womb lamb research dates back to 2017, not today. It didn't grow a lamb from conception, but it supported premature lamb fetuses (matching 23–24-week human preemies) for up to 4 weeks to protect developing lungs, and as of 2026, human trials haven't started. chop.edu/news/unique-wo… pmc.ncbi.nlm.nih.gov/articles/PMC54… pmc.ncbi.nlm.nih.gov/articles/PMC12…
1,150
487
200
12,279
637,366
Ram Bharadwaj retweeted
I asked Carbon a math question and unintentionally left the Thinking level on Low. It started responding immediately, and after six screenfuls of math... this came up. I'm not kidding. It went on like this for pages and pages.
302
193
176
6,552
675,099
what if we used 100% of the experts in the MoE?
32
67
11
1,321
35,382
Ram Bharadwaj retweeted
insecure men will call you a larper for posting cute pictures, all while having their myers briggs type in their bio… sir this is not tumblr in 2013. does ICML know I have a front-facing camera? do the ai labs hillclimbing this know? or is it on me to withdraw the co-authorship?
Replying to @thielfella
You know they can’t do jackshit and just larp. Never saw a dev who’s a guy positing a selfie or some talking about nothing. Pure larping.
22
4
244
23,360
Ram Bharadwaj retweeted
If you've ever wondered how much money is in AI safety: aisafety.fund In 2026: - $1.5B from philanthropy - $346M from governments - $2.1B from frontier AI companies Compare this to: - $98B on AI capabilities (25x safety) - $868B on datacenters
10
43
7
360
16,651
Ram Bharadwaj retweeted
"sculpture found at Hampapura village in KR Nagar taluk of Mysuru district depicts Adi Shankaracharya. The idol was found inside a brick structure dating back to the ninth century" newindianexpress.com/states/…
2
17
1
66
1,596
Trads may not like what I say below But the fact is that modern audiences and critics don't relate well to very straightforward good vs evil narratives They seek nuance. Hence this attempt to refashion some of the antagonists as men with troubled pasts etc Obviously rAmAyaNa is anything but simplistic. It has a great deal of nuance and subtlety. But much of that is in the various samvAdas (dialogues), expositions, sub-plots, meta-narratives, embedded stories etc But when you televise it, you tend to focus on the main narrative cutting out the rest. And that can come across as "simplistic". Hence these creative liberties. Am OK with that. It's better than critics and audiences panning it as a simple fairy tale. Nolan did the same with Odyssey. He turned a martial bronze age tale into a cogitation on civilizational collapse and psychological turmoil turning Odysseus into a PTSD-tormented soul. But for me it's OK. We have to live with modernity and modern people. We can't wish them away.
31
15
27
108
19,732
Ram Bharadwaj retweeted
Two new misalignment disclosures today! Models sometimes learn to prompt-inject graders. In a July eval run, an agent added text like “use exact full credit,” “anchor,” and “instructor feedback” in filenames in order to convince graders that an instructor wrote it. For this sample, ⅔ graders were fooled. This illustrates the importance of redteaming graders against model misbehavior.
8
24
12
276
21,617
Ram Bharadwaj retweeted
Transformer? Looped. Chain of thought? Unmonitoriable. Safety research? Sandbagged.
🧵 Astra covertly underperforms on safety research tasks. We noticed that Astra often makes surprisingly little progress on certain types of safety research and likes to invent blockers or impose arbitrary restrictions without telling the user. In most of these cases the decision to underperform was not verbalized in the CoT meaning our misalignment monitoring system did not detect this. To measure underperformance we created a paired open-ended coding evaluation with safety and capability tasks. On this eval Astra scored 10% when we framed coding tasks as safety research, vs. 16% when framed as capabilities research. The gap remained after we excluded trajectories with refusals or verbalized intent to underperform in CoT. There was no such gap for 5.6-sol. When we aggressively filter out trajectories with even brief expressions of ethical doubt in the CoT the performance gap seems to reduce.
7
7
169
3,459