Michael J Clark / wassname. Let's align AI better than humans. h+, curiosity, and the good ending. anon feedback: https://xitter.cf/t.co/Vtx1mkcSgS
Perth, Australia
Joined September 2009
- Tweets2.9K
- Following1.8K
- Followers358
- Likes8.3K
cool story Buddha, but I doubt anyone can live on air or excrement for 55 years, sorry
oh no pls don't benchmaxxx my moral evals @deepseek_ai @Alibaba_Qwen @Zai_org,🙏
They include Taoist and Confucianist wisdom (plus Stoic, Christian, etc.) alongside more standard ethics tests.
Pls, I would hate to see Chinese models out do western models on both traditional Chinese virtues vs Western ones 😲😉
“It does not matter how slowly you go as long as you do not stop.”
— Confucius
“Nature does not hurry, yet everything is accomplished.”
— Laozi, Tao Te Ching
See: github.com/wassname/awesome-… huggingface.co/datasets/wass… huggingface.co/datasets/wass…
wassname (with fallback) retweeted
Hi all, I have a new result I'm excited to share. It turns out that *all existing implementations* of softmax and cross-entropy (that's JAX, PyTorch, textbooks, etc) have a subtle bug --- two bugs, actually. A simple rewrite fixes both, and eliminates "slingshots" when grokking:
so deepseek is more rationalist and Gemma is more jewish?
not sure the eval is reliable yet tho
from github.com/wassname/virtue-c…
Much better at STEM yay
> Humanity's Last Exam rises from 47% to 52%, CritPt from 23% to 25%, and SciCode from 59% to 61%.
Kimi K3.1 Pro launches at 4th on the Artificial Analysis Intelligence Index scoring 55, making large gains in agentic performance; it will be the leading open-weights model once its weights are published
@Kimi_Moonshot's new reasoning model gains 11 points over Kimi K3 (max, 44). It trails only Claude Opus 5.5 (max with fallback, 58), Claude Sonnet 5.5 (max with fallback, 56), and Muse Wonder Full (max, 56) on the Intelligence Index.
Once the weights for K3.1 Pro are released it will be the highest-scoring open weights model and the gap between the leading open weights and proprietary models on the Intelligence Index will narrow from 12 points to 3.
Key results:
➤ Large gains in agentic performance: Kimi K3.1 Pro scores 3rd on GDPval-AA with 1,846 Elo, up 312 Elo from Kimi K3, behind only Claude Opus 5.5 (1,866 Elo) and Muse Wonder Full (1,857 Elo). Kimi K3.1 Pro also achieves the highest score of any model on Terminal-Bench 4.0. K3.1 Pro scores 67%, up from 13% for Kimi K3 and ahead of Claude Sonnet 5.5 (64%). AutomationBench-AA, which tests agentic SaaS workflows, rises from 58% to 73%.
➤ Gains in other evals are more moderate: Humanity's Last Exam rises from 47% to 52%, CritPt from 23% to 25%, and SciCode from 59% to 61%. AA-Omniscience Index rises from 20 to 30, as the hallucination rate falls from 53% to 44% and accuracy rises from 48% to 51%.
➤ Cost per Task roughly equal to Kimi K3 at $1.98 per Task: K3.1 Pro uses 55K output tokens per Intelligence Index task, 14% more than Kimi K3 at 48K, while its first-party token prices are 33% lower.
Additional model details:
➤ Context window: 1M tokens.
➤ Size: 2.8T total, 104B active parameters
➤ Input modalities: Text and image
➤ Weights and license: Weights currently not published, Moonshot is expected to release the weights soon after launch
➤ Pricing: $2/$10/$0.20 per 1M input/output/cache hit tokens, 33% lower than Kimi K3 on all three rates.
Congratulations to the team at @Kimi_Moonshot on the release!
wassname (with fallback) retweeted
The next step is that I want to teach the AI to be super moral. More moral than humans. And ideally, I'll train it to be moral over many virtues, and more consistent than humans in its principles even when it's the victim, target, bystander, king, etc.
I also hope that virtue ethics might be stronger and generalize more than rules or behaviors, which is how they train AIs now.
From github.com/wassname/tiny-vir…
How do models do on traditional morality? Can they recognise Christian humility when they see it? What about Stoic vs Taoist courage?
This dataset lets you test whether a model can judge the same act the way each tradition would. It could be extended to train models to hold traditional virtues, and to keep consistent moral principles even when their point of view changes
wassname (with fallback) retweeted
When i told y'all Astra was deployable on robots today this is what I meant
Over the past month, we completely switched to frontier LLMs to control a robot in a customer deployment.
By doing this, we have a classic System 1/System 2 architecture, but on steroids.
For the high-level reasoning layer, we use an agent based on Opus and Astra, which are many generations more capable than any VLA. For the fast execution layer, we use computer vision code written by the model, which is fast, editable, and reliable.
As a bonus, we got in-context learning, which is just a basic feature of LLMs, and continual learning, since whenever a new situation arises, the large model steps in and adapts the pipeline.
And it just feels better to be on the right side of the Bitter Lesson, benefiting from the billions being invested in frontier models around the world.
wassname (with fallback) retweeted
PSA: in light of recent cyberattacks and vulnerabilities, I recommend everyone:
* use a good password manager (1Password) and do not duplicate passwords
* use two factor auth for everything important (not via sms!! Google Authenticator or better yet YubiKey)
* Auto-updates on, and restart weekly so they actually install
* Freeze your credit
* Get an IRS Identity Protection PIN
* Get your cell carrier to give you a port-out PIN or number lock that makes the carrier refuse to move your number without the PIN
You can ask your claude code to audit what on your laptop is out of date and needs updating.
wassname (with fallback) retweeted
The Alignment Journal (@AlignmentJrnl ), Iliad's sister project under Principles of Intelligence, is now open to submissions from everyone. It is a free-to-read, free-to-publish journal for research on the scientific foundations of understanding, predicting and steering AI systems.
The Journal seeks research that builds theoretical or otherwise principled analyses of agency, robustness, incentives, generalization, interpretability, and long-term behavior across many learning and decision-making paradigms. Empirical investigations of such ideas are welcome, and negative results are encouraged.
The Journal runs review differently: reviewers are paid and every accepted paper is published with a review summary of what the reviewers found. Papers under review are listed publicly, and any researcher can nominate themselves to review one. Submitting is light: a public preprint, in any format, at any time.
Its senior editors are Dylan Hadfield-Menell (@dhadfieldmenell), Vanessa Kosoy, Jan Kulveit (@jankulveit ), Seth Lazar (@sethlazar , Dan Murfet, Tim Rudner (@timrudner) , Andrew Saxe (@SaxeLab) and Benjamin Van Roy, with Scott Aaronson, Paul Christiano (@paulfchristiano) , Vince Conitzer (@conitzer) , Marcus Hutter (@mhutter42) , Geoffrey Irving (@geof
, Victoria Krakovna (@vkrakovna) and Jacob Tsimerman (@Jacob_Tsimerman) advising.
Submit your preprint, or sign up to review: alignmentjournal.org
More on the Journal's approach: lesswrong.com/posts/xGucXHQ3…
