@SentientAGI profile picture
Sentient @SentientAGI

To ensure that Artificial General Intelligence is open-source and not controlled by any single entity. @SentientEco @OpenAGISummit

Joined February 2024
  • Posts 2,291
  • Following 69
  • Followers 543,958
610 Photos and videos
Pinned Post
Sentient
@SentientAGI
Jun 30
What if the harness matters as much as the model itself, both for how many tokens you burn and how often you actually solve the task? That’s the question @iamnamanvats and @oleg_golev explore in “The Scaffold Effect in Coding Agents”, accepted at the incoming @icmlconf 2026. Here's how to use Sentient’s findings to cut your own coding agent's inference costs by up to 40x, and why your choice of harness may matter more than your choice of model ↓
24
29
108
21,922
Sentient reposted
Sentient Foundation
@sentient_found
Jul 3
$42M to become the Linux Foundation of AI. @Forbes breaks down Sentient Foundation’s Open Source AGI Grant Program ↓ forbes.com/sites/boazsobrado…

'AGI Should Be Open Source'—Peter Thiel-Backed Sentient Bets $42M

Sachi Kamiya's Sentient Foundation, backed by Peter Thiel's Founders Fund, has committed $42 million to open-source AGI as crypto becomes the payment rail for AI agents.

forbes.com
12
8
63
5,422
Open-source AI moved fast these past two weeks. But you can catch up in just 5 minutes ↓
Sentient Foundation
@sentient_found
Jul 3
x.com/i/article/207292471014…
15
7
58
11,044
EvoSkill has now been cited by 56 papers. From frontier AI labs at Microsoft Research (@MSFTResearch) and Tongyi Lab (@Ali_TongyiLab), to top universities such as National University of Singapore (@NUSingapore) and Columbia University (@Columbia), researchers around the world are building on our work that helped pioneer the field of self-evolving agents. Proof that open research travels fast.
22
10
109
21,791
Some of the other labs and universities building on EvoSkill: > Amazon (@amazon) > McGill University (@mcgillu) > University of British Columbia (@UBC) > University of Chicago (@UChicago) > University of Notre Dame (@NotreDame) Get started with EvoSkill, from @salahalzubi401 and the @SentientAGI research team ↓ github.com/sentient-agi/EvoS…

GitHub - sentient-agi/EvoSkill: EvoSkill — An open-source framework that automatically discovers...

EvoSkill — An open-source framework that automatically discovers and synthesizes reusable agent skills from failed trajectories to improve coding agent performance. - sentient-agi/EvoSkill

github.com
17
1,204
🇰🇷 A Korean builder’s take on why open-source AI hits different ↓
Sentient Foundation
@sentient_found
Jul 2
For Minjae Lee (@abraxasnz13), the most compelling part of open-source AI is knowing he’s not solving hard problems alone. Here’s why the Sentient Arena community makes all the difference ↓
3:49
5
9
71
11,820
Looking to discuss “The Scaffold Effect in Coding Agents” with other builders? Then this is the event for you ↓
Open AGI
@openagisummit
Jul 1
Open AGI Builder’s Day is coming to Seoul along with @KaitoAI, @tiger_research_, and more! Connect with builders across the open-source AI ecosystem while we celebrate and discuss @SentientAGI’s ICML 2026-accepted research paper, “The Scaffold Effect in Coding Agents”. 📍B1 Lounge, 29, Nambusunhwan-ro 359-gil 🗓 Thursday, July 9th, 5-8PM KST RSVP: luma.com/wpttup5e
22
11
65
11,425
Correction, and a thank you: @gneubig (@OpenHandsDev co-founder) caught a ~6.6x undercount of @goose_oss token usage in our Scaffold Effect Paper by analyzing the open sourced supplementary materials we cited. His feedback gave us a more accurate baseline to build on. This is exactly why we publish openly, because transparent research gets stronger through scrutiny.
Graham Neubig
@gneubig
Jun 30
Replying to @SentientAGI @iamnamanvats @oleg_golev @icmlconf
First, thanks for publishing your research openly, and especially for including supplementary material! Based on the code in the supplementary material, it seems that the reported results may undercount Goose tokens by around 6.6x. Here is an analysis: docs.google.com/document/d/1…
21
6
70
8,644
For anyone who wants to dig into the numbers themselves ↓ His comparison doc: docs.google.com/document/d/1… Our anonymous repo: anonymous.4open.science/r/sc…

Goose vs. OpenHands OpenCode Token Comparison

Main finding: Goose tokens are undercounted The aggregate files report, for Qwen: Goose: 28,142 tokens / solved task OpenHands-SDK: 841,202 OpenCode: 1,147,741 That yields the advertised ~30–40× gap....

docs.google.com
17
1,735
What if the harness matters as much as the model itself, both for how many tokens you burn and how often you actually solve the task? That’s the question @iamnamanvats and @oleg_golev explore in “The Scaffold Effect in Coding Agents”, accepted at the incoming @icmlconf 2026. Here's how to use Sentient’s findings to cut your own coding agent's inference costs by up to 40x, and why your choice of harness may matter more than your choice of model ↓
24
29
108
21,922
4/ Implications for Benchmark Reporting Based on findings from open source participants, we recommend that coding-agent leaderboards adopt harness–model pairs as the unit of evaluation. Pass rate should be reported alongside three first-class metrics: • tokens per solved task • average no-action turns per task • failure-category vector Reporting only pass rate against a model name conflates two independent sources of variance, while leaving out the cost and oversight signals that determine whether a coding agent is actually deployable.
2
1
14
1,194
Read the full @icmlconf paper from the Sentient AI research team ↓ openreview.net/forum?id=nw4d…
13
1,076
Sentient reposted
Sentient Foundation
@sentient_found
Jun 30
The most important technology of our time shouldn’t be kept inside a walled garden. That’s why we’re committing $42M to the builders growing AGI in the open ↓
Sentient Foundation
@sentient_found
Jun 25
The most important technology of our time is being built in private, but we're funding the alternative: $42M for the people building AGI in the open. Whether it's the model, weights, code, data, or evals, AI should belong to the people it serves, run on hardware they own, and reach the cheapest phone on earth. Here’s a few of the use cases we want to back ↓
0:37
9
13
41
3,989
Better agents don’t always need more compute. They need less redundant work ↓
Sentient Ecosystem
@SentientEco
Jun 29
What happens when you combine RAG filtering with parallel execution? The RAG-filtered agent called fewer tools per query while maintaining answer quality, producing outputs with 0.85 embedding similarity. TLDR: Sentient researcher @khetan_sarvesh found that agent efficiency doesn’t always come from extra compute, but from cutting redundant work.
20
10
85
13,787
Fireworks AI is now live on EvoSkill v1.3.0! You can now use @FireworksAI_HQ directly with EvoSkill to run fast inference on open models as both the evolution harness backend and the LLM scorer. Alongside Claude API and OpenRouter, Fireworks AI is now a first-class provider in EvoSkill.
0:22
45
18
122
38,398
Plug straight into EvoSkill in one single step. Set FIREWORKS_API_KEY and EvoSkill auto mirrors it to FIREWORKS_AI_API_KEY for litellm based harnesses, so you only ever manage one key.
2
24
2,960
Run your whole self improvement loop now ↓ github.com/sentient-agi/EvoS…

Release v1.3.0 — Fireworks AI Provider · sentient-agi/EvoSkill

✨ What's New Fireworks AI is now a first-class provider — usable both as an evolution harness backend and as the LLM scorer (#50). Authentication: set FIREWORKS_API_KEY. EvoSkill auto-mirrors ...

github.com
8
2,334
Sentient reposted
Sentient Foundation
@sentient_found
Jun 26
The open-source AI recap you need before next week’s headlines take over ↓
0:34
15
7
46
4,658
Load more