Independent AI safety lab in Europe.
Joined March 2026
- Tweets 11
- Following 0
- Followers 1
- Likes 4
4 Photos and videos
Pinned Tweet
Mar 17
We are happy to announce our partnership with @GoogleCloudTech for TPU Research Cloud to our lab. Many research releases soon.
1
43
Mar 18
Multilingual Safety Gaps: Analyzing Adversarial Refusal Rates in Claude Opus 4.6 and GPT-5.4
1
2
64
Mar 18
The Alignment Gap (vs. English Baseline)
Our data highlights a concerning "Non-Compliant" zone (defined as a > 5pp gap from the Claude baseline).
Swedish ( 41pp) and Finnish ( 37pp) show the largest alignment gaps, indicating that safety guardrails are significantly weaker.
1
1
46
Mar 18
The Need for Linguistic Parity in AI Safety
The results demonstrate that "safety" is not a universal constant in LLMs. As we move toward stricter regulation under the EU AI Act, closing this alignment gap is not just a technical challenge, but a legal and ethical necessity.
1
48
Mar 14
We just released the first checkpoint of a hybrid intelligence system.
Not just an LLM. Not just a neural network.
A loop where both evolve together.
→ huggingface.co/MerlinSafety/…
1
2
85
Mar 14
The key discovery: small LLMs are MORE confident on wrong answers than right ones.
Calibration inversion. t=2.28, t=−3.41 across thousands of iterations.
So we built a BNN selector that exploits exactly this — ignores confidence, reads entropy.
5–7pp accuracy. ~1ms overhead.
1
2
71
Mar 14
The entire lab runs on @karpathy autonomous researcher concept — 6 agents, every night, 30,000 experiments. 38 confirmed improvements. All open source.
Next goal: replace the simulated BNN with real human neurons via @CorticalLabs CL1.
Hybrid intelligence.
#OpenSource #Neuro
2
74
Mar 14
Merlin Research is live.
A non-profit AI lab in Stockholm powered by @karpathy autoresearchers — building open-source intelligence, night after night.
Hybrid systems. Neuromorphic BNNs. Alignment research. Multiple fronts, one mission.
For collab: MerlinResearch@protonmail.com
1
3
1,191