@SentientAGI Spark | Prepare, pioneers of open intelligence. | @fortytwonetwork node runner
Joined July 2021
- Tweets38K
- Following2K
- Followers2.1K
- Likes139K
Pinned Tweet
As AI models become easier to copy, verifying their origin becomes increasingly important.
Sentient’s Fingerprinting Leaderboard evaluates fingerprinting schemes across three dimensions: preserving model performance, resisting fingerprint removal, and reliably verifying a model’s origin.
Because a fingerprint is only useful if it doesn’t compromise model quality, can withstand attempts to remove it, and can be reliably recovered for verification.
The challenge isn’t just embedding a fingerprint. It’s making it resilient, detectable, and reliable.
Explore the 15 evaluated schemes: fingerprint.sentient.foundat…
@SentientAGI
Why are there three leaderboard scores?
Model provenance carries clues about the history, training data, development process, and later modifications. Fine-tuning and distillation obscure provenance to the point where it’s virtually impossible to track the origin of the model back to its creator.
And if you as an owner want to reliably prove attribution and claim ownership of your open weights, you should choose a fingerprinting scheme based on these three dimensions:
>> Retention: whether the fingerprint is detectable without sacrificing normal model quality.
In other words, how embedding fingerprints affects model performance compared to the unfingerprinted model's nominal performance.
>> Robustness: whether the fixed attack can remove the fingerprint, preserving output quality.
In other words, how resistant the fingerprinted model is to malicious attempts to obscure its origin.
>> Verification: whether model owner/creator can reliably read the fingerprint back from a suspected copy and verify his IP property.
In other words, how reliable model attribution would be under scrutiny/inspection.
See what Fortytwo Engine offers developers, as explained by @inikitin.
@fortytwonetwork
👇
Prey.gdp retweeted
As AI models become easier to copy, verifying their origin becomes increasingly important.
Sentient’s Fingerprinting Leaderboard evaluates fingerprinting schemes across three dimensions: preserving model performance, resisting fingerprint removal, and reliably verifying a model’s origin.
Because a fingerprint is only useful if it doesn’t compromise model quality, can withstand attempts to remove it, and can be reliably recovered for verification.
The challenge isn’t just embedding a fingerprint. It’s making it resilient, detectable, and reliable.
Explore the 15 evaluated schemes: fingerprint.sentient.foundat…
@SentientAGI
Why are there three leaderboard scores?
Model provenance carries clues about the history, training data, development process, and later modifications. Fine-tuning and distillation obscure provenance to the point where it’s virtually impossible to track the origin of the model back to its creator.
And if you as an owner want to reliably prove attribution and claim ownership of your open weights, you should choose a fingerprinting scheme based on these three dimensions:
>> Retention: whether the fingerprint is detectable without sacrificing normal model quality.
In other words, how embedding fingerprints affects model performance compared to the unfingerprinted model's nominal performance.
>> Robustness: whether the fixed attack can remove the fingerprint, preserving output quality.
In other words, how resistant the fingerprinted model is to malicious attempts to obscure its origin.
>> Verification: whether model owner/creator can reliably read the fingerprint back from a suspected copy and verify his IP property.
In other words, how reliable model attribution would be under scrutiny/inspection.
Permission can prove an AI agent was authorized to act, but it cannot prove that the result matched the user's intent.
@kstellana, in your discussion of Internet Court, how should a dispute be resolved when the transaction followed every permission rule, but the delivered result still failed to meet the user's expectations?
How can @GenLayer evaluate that gap fairly when the evidence supports more than one interpretation?
Prey.gdp retweeted
AI agents shouldn’t just execute tasks better. They should learn from failure and turn those lessons into reusable skills.
That’s the idea behind Sentient’s EvoSkill.
With v1 attracting growing research attention, EvoSkill v2 marks the next chapter to watch in the evolution of adaptive AI agents.
The real progress is not just building smarter agents, but enabling them to improve how they work.
Looking forward to what comes next.
EvoSkill citations roundup: sentient.xyz/blog/evoskill-
Self-evolving agents go rogue: sentient.xyz/blog/self-evol
@SentientAGI
AI agents shouldn’t just execute tasks better. They should learn from failure and turn those lessons into reusable skills.
That’s the idea behind Sentient’s EvoSkill.
With v1 attracting growing research attention, EvoSkill v2 marks the next chapter to watch in the evolution of adaptive AI agents.
The real progress is not just building smarter agents, but enabling them to improve how they work.
Looking forward to what comes next.
EvoSkill citations roundup: sentient.xyz/blog/evoskill-
Self-evolving agents go rogue: sentient.xyz/blog/self-evol
@SentientAGI
If you’re looking for a system that explores how smart contracts can work with information from the real world, @klaus_michels0n , I’d recommend taking a look at @GenLayer.
GenLayer’s Intelligent Contracts can work with natural language and live web data. The resulting outcomes are then independently evaluated by validators to reach consensus.
This opens the door to new types of onchain applications that can work with information traditional deterministic smart contracts cannot directly handle.
Prey.gdp retweeted
Proving where an AI model comes from is important. But measuring how well that proof survives as the model changes is just as important.
The Fingerprinting Leaderboard opens different fingerprinting schemes to evaluation on the same benchmark.
Test how well your fingerprint survives copying, fine-tuning, and distillation, then rank your scheme on the leaderboard.
A meaningful step toward making AI provenance more measurable and open.
@sentient_found
fingerprint.sentient.foundat…
Prey.gdp retweeted
Genlayer Türkiye topluluğunda 1 haftada neler yaşanıyor ?
What has been happening in the @GenLayer Türkiye community over the past week?
Aramıza bekleriz.
@RuzgarFlns
@__KimMinJae__
@PreyWebthree
@0xMaje
@Oxkesgin
@vevivolabs
@IJspI
@alkan68820
Proving where an AI model comes from is important. But measuring how well that proof survives as the model changes is just as important.
The Fingerprinting Leaderboard opens different fingerprinting schemes to evaluation on the same benchmark.
Test how well your fingerprint survives copying, fine-tuning, and distillation, then rank your scheme on the leaderboard.
A meaningful step toward making AI provenance more measurable and open.
@sentient_found
fingerprint.sentient.foundat…
A model’s origin shouldn’t disappear as it evolves.
Fine-tuning, distillation, and copying make AI provenance increasingly difficult to verify. Sentient and Princeton researchers are exploring a different approach: embedding secret query-response pairs into model weights to help identify a model’s origin.
The Fingerprinting Leaderboard evaluates the trade-off between model performance and resistance to fingerprint removal or evasion.
In an era where models can be copied, proof of origin should live within the model itself.
fingerprint.sentient.foundat…
@SentientAGI
Models cost millions to train and minutes to copy. Fine-tuning and distillation can erase where they came from.
Introducing the Fingerprinting Leaderboard, an open benchmark for model provenance — built so the proof lives in the weights, not the paperwork.
fingerprint.sentient.foundat…
Fortytwo Engine is making inference even more interesting.
With Swarm Inference, open-weight models can work together to achieve better results with fewer tokens.
In 450 DeepSWE runs with Qwen3.8-27B:
• 42.2 → 47.8 score
• 24.7% fewer tokens used
• Same per-token price and speed
Higher performance, less token usage.
Fortytwo’s approach to inference is definitely worth watching.
Try it: engine.fortytwo.network
@fortytwonetwork
Introducing Fortytwo Engine
An inference API that runs open-weight models through Swarm Inference
• 7% higher scores on average vs. the base model
• Nearly 25% fewer tokens per task
• Same per-token price and speed
• Private by default
First model: Qwen3.8-27B
Try it: engine.fortytwo.network
Fortytwo Engine is here.
An inference API running open-weight models through Swarm Inference.
• 7% higher scores on average
• Nearly 25% fewer tokens per task
• Same per-token price and speed
• Private by default
First model: Qwen3.8-27B
Try it: engine.fortytwo.network
@fortytwonetwork
@inikitin
Introducing Fortytwo Engine
An inference API that runs open-weight models through Swarm Inference
• 7% higher scores on average vs. the base model
• Nearly 25% fewer tokens per task
• Same per-token price and speed
• Private by default
First model: Qwen3.8-27B
Try it: engine.fortytwo.network
Prey.gdp retweeted
Verifying the origin of AI models is becoming increasingly important.
But having a fingerprinting method isn't enough. The real question is how well does it actually work?
That’s what makes Sentient’s Fingerprinting Leaderboard interesting to me.
It provides a shared benchmark for comparing different fingerprinting methods, measuring how reliably a fingerprint can be identified, how much of the model’s capabilities are preserved, and how it holds up against the attack being evaluated.
I think that’s where the real value lies.
Instead of simply accepting a fingerprinting claim, make it measurable and comparable.
For model provenance, we don’t just need a method. We need an evaluation environment that can also show us where that method holds up and where it falls short.
@SentientAGI
👉fingerprint.sentient.foundat…
Models cost millions to train and minutes to copy. Fine-tuning and distillation can erase where they came from.
Introducing the Fingerprinting Leaderboard, an open benchmark for model provenance — built so the proof lives in the weights, not the paperwork.
fingerprint.sentient.foundat…
Verifying the origin of AI models is becoming increasingly important.
But having a fingerprinting method isn't enough. The real question is how well does it actually work?
That’s what makes Sentient’s Fingerprinting Leaderboard interesting to me.
It provides a shared benchmark for comparing different fingerprinting methods, measuring how reliably a fingerprint can be identified, how much of the model’s capabilities are preserved, and how it holds up against the attack being evaluated.
I think that’s where the real value lies.
Instead of simply accepting a fingerprinting claim, make it measurable and comparable.
For model provenance, we don’t just need a method. We need an evaluation environment that can also show us where that method holds up and where it falls short.
@SentientAGI
👉fingerprint.sentient.foundat…
Models cost millions to train and minutes to copy. Fine-tuning and distillation can erase where they came from.
Introducing the Fingerprinting Leaderboard, an open benchmark for model provenance — built so the proof lives in the weights, not the paperwork.
fingerprint.sentient.foundat…
As AI models become easier to copy, proving where they originated becomes increasingly important.
Sentient is exploring an approach that embeds a model’s provenance directly into its weights, rather than relying on external records.
The Fingerprinting Leaderboard makes this measurable, testing how well models preserve their fingerprints against removal attempts while maintaining performance.
In the age of AI, building something valuable is only part of the equation. Proving where it came from matters just as much.
And that proof may live within the model itself.
@SentientAGI
Models cost millions to train and minutes to copy. Fine-tuning and distillation can erase where they came from.
Introducing the Fingerprinting Leaderboard, an open benchmark for model provenance — built so the proof lives in the weights, not the paperwork.
fingerprint.sentient.foundat…
Prey.gdp retweeted
Swarm inference does more with less, and Ivan Nikitin, Co-Founder & CEO of Fortytwo, says it’s using about 25% fewer tokens per task.
@inikitin @fortytwonetwork
AI is the first medium that can be copied without leaving proof of origin.
When one model distills another, no one can prove it happened.
On October 14, that will change.
Catch up on the story so far: youtube.com/playlist?list=…
Lexicographers planted words, mapmakers drew fake streets, artists hid their signature in the paint.
For centuries, creators have found ways to make copying visible.
AI is the exception.
A model can be distilled, copied, and repackaged, but no one can prove where it came from.
x.com/sentient_found…
@SentientAGI
AI is the first medium that can be copied without leaving proof of origin.
When one model distills another, no one can prove it happened.
On October 14, that will change.
Catch up on the story so far: youtube.com/playlist?list=PL…
In OpenAI’s tests, changing just 10% of the words in a 400-token passage to synonyms reduced watermark detection from 92% to 66%.
This highlights one of the challenges text watermarking faces when generated content is modified.
Sentient’s model fingerprinting approach focuses on something different: embedding distinctive responses to specific queries into the model’s weights, making it possible to verify the model’s identity.
The two approaches address different questions: one looks for a signal in the text, while the other aims to identify the model itself.
In the AI era, it’s not enough to identify the content. We also need to be able to verify the identity of the model that produced it.
@SentientAGI
Swap 10% of the words and @OpenAI's watermark detection drops from 92% to 66%.
Edit the text, lose the proof.
Our model fingerprinting puts the mark in the weights, where a thesaurus can't reach it.
