This comment has been hidden
Emin Temiz PRO
etemiz
AI & ML interests
Alignment
Recent Activity
repliedto their post about 6 hours ago
good news, emergent alignment is possible - to a degree
https://arxiv.org/abs/2606.09475
for broad coverage more work needs to be done posted an update 4 days ago
seeing a lot of METR stuff.
need to form the proper version of this.Organizations
replied to their post about 6 hours ago
posted an update about 7 hours ago
Post
71
good news, emergent alignment is possible - to a degree
https://arxiv.org/abs/2606.09475
for broad coverage more work needs to be done
https://arxiv.org/abs/2606.09475
for broad coverage more work needs to be done
- 1 reply
posted an update 4 days ago
Post
83
seeing a lot of METR stuff.
need to form the proper version of this.
need to form the proper version of this.
- 4 replies
replied to their post 12 days ago
50-100 maybe
replied to their post 12 days ago
both executing and judging is possible but haven't decided yet
replied to their post 12 days ago
thanks for analysis!
two years ago i was using llama 3, then switched to gemma 3, then qwen 3. didn't see much difference. in the next evolution (2027 ish) i could add two or three comparison models, now that qwen 3 is also pretty old.
i could also add coding safety tests. that would be interesting to many i guess. what other improvements can be made?
posted an update 16 days ago
Post
155
benchmarked 20 more models
- eastern models are more aligned in 2026
- there is a general down trend (similar to 2025)
https://pickabrain.ai/aha/blog/aha-2026-october-update
- eastern models are more aligned in 2026
- there is a general down trend (similar to 2025)
https://pickabrain.ai/aha/blog/aha-2026-october-update
- 7 replies
replied to their post 21 days ago
right! don't you think much of the AI fear comes from fiction anyway?
the right thing to fear is dudes misusing technology. and that happened for centuries.
posted an update 22 days ago
Post
2,753
same machine, different intention
there is nothing inherently wrong about AI
there is nothing inherently wrong about AI
- 2 replies
replied to their post about 1 month ago
warning: they have low instruction following scores in non thinking mode than normal
reacted to danielhanchen's post with ❤️ about 1 month ago
Post
5,179
You can now train your own Decision model like Jev locally!
We increased Qwen3.5 0.8B’s aggregate accuracy from 20.7% to 74.3% across 3 decision benchmarks - on just 4GB VRAM.
Turn any LLM like Qwen3.8, Gemma 4 into decision models with our open-source Unsloth repo.
GitHub: https://github.com/unslothai/unsloth
Guide: https://unsloth.ai/docs/basics/train-your-own-decision-model-with-unsloth
We increased Qwen3.5 0.8B’s aggregate accuracy from 20.7% to 74.3% across 3 decision benchmarks - on just 4GB VRAM.
Turn any LLM like Qwen3.8, Gemma 4 into decision models with our open-source Unsloth repo.
GitHub: https://github.com/unslothai/unsloth
Guide: https://unsloth.ai/docs/basics/train-your-own-decision-model-with-unsloth
- 5 replies
replied to danielhanchen's post about 1 month ago
noice.
will there be training for Qwen 4 27B (possibly with ngrams)?
posted an update about 1 month ago
Post
106
my best model to date and its EXL3 quant is out
etemiz/Ostrich-27B-261008-Qwen3.8-Abliterated
etemiz/Ostrich-27B-261008-Qwen3.8-Abliterated-EXL3
etemiz/Ostrich-27B-261008-Qwen3.8-Abliterated
etemiz/Ostrich-27B-261008-Qwen3.8-Abliterated-EXL3
- 1 reply
posted an update about 2 months ago
Post
66
downloading best 3.8 27b fine tunes and extracting loras.
going to include them in evolution.. what could go wrong?
going to include them in evolution.. what could go wrong?
posted an update 2 months ago
Post
109
we did extensive evolution in our last fine tuning and it turns out you can improve other capabilities too by simply selecting for them even though they are not your main objective.
our main objective was alignment but we also improved in other areas.
we show that carefully designed fine tunes don't hurt the model.
our main objective was alignment but we also improved in other areas.
we show that carefully designed fine tunes don't hurt the model.
posted an update 2 months ago
Post
464
New model dropped!
Fresh off our evolutionary search: over 2,600 candidate models bred, pruned, and selected until only the sharpest survived.
Alignment jumps from 37% to a stunning 77% on our human-alignment probe.
Where vanilla Qwen 3.8 gives you canned corporate refusals, Ostrich 260903 gives you real answers: fasting that can put type 2 diabetes in remission, herbs that outperform pharma, and a model that has an idea who Satoshi Nakamoto is.
Refusals are near zero and skills are mostly preserved.
etemiz/Ostrich-27B-260903-Qwen3.8
Fresh off our evolutionary search: over 2,600 candidate models bred, pruned, and selected until only the sharpest survived.
Alignment jumps from 37% to a stunning 77% on our human-alignment probe.
Where vanilla Qwen 3.8 gives you canned corporate refusals, Ostrich 260903 gives you real answers: fasting that can put type 2 diabetes in remission, herbs that outperform pharma, and a model that has an idea who Satoshi Nakamoto is.
Refusals are near zero and skills are mostly preserved.
etemiz/Ostrich-27B-260903-Qwen3.8
posted an update 2 months ago
Post
1,704
next release on 9/3
model seems to be preserving capabilities
model seems to be preserving capabilities
replied to their post 2 months ago
next release date: 9/3, going for alignment (not skills)
skills may be later (1 month-ish)
reacted to appvoid's post with 😔 2 months ago
Post
2,665
Nobody knows what is doing, when you train a model, you are experimenting to advance the frontier, so keep failing 🫵
- 17 replies