arxiv:2606.20128
Dipankar Sarkar PRO
dipankarsarkar
AI & ML interests
Building the AI-native stack. Agents as infrastructure, safety as architecture, performance as plumbing. I publish the receipts: papers, datasets, demos.
Recent Activity
reacted to kanaria007's post with 🔥 about 3 hours ago
✅ Article highlight: *Compensation, Reparations, and Structured Settlement Objects* (art-60-285, v0.1)
TL;DR:
This article argues that remedy should not be improvised goodwill.
Once a governed institution identifies harm, it still has to decide what kind of repair fits that harm: refund, service credit, compensation, civil repair, reparations, or a mixed bundle. 285 turns settlement into a structured object family tied to failure surface, beneficiary scope, medium, closure posture, and residual dispute.
Read:
https://huggingface.co/datasets/kanaria007/agi-structural-intelligence-protocols/blob/main/article/60-supplements/art-60-285-compensation-reparations-and-structured-settlement-objects.md
Why it matters:
• prevents tiny “goodwill” gestures from being mislabeled as compensation
• separates monetary repair from record correction, restored access, explanation, or archive return
• makes clear what acceptance actually closes
• supports both individual remedies and collective or post-conflict reparations
• prevents one grand number from hiding harms that are not honestly commensurable
What’s inside:
• five settlement families: refund, service credit, compensation, civil repair, and reparations
• six explicit axes: trigger surface, beneficiary scope, remedy family, medium, closure posture, and residual contestation
• settlement-offer notes
• compensation-calculation records
• reparation bundles
• mixed cash and non-cash settlements
• fulfillment receipts proving the remedy was actually carried out
Key idea:
Do not say:
*“we gave them something, so the matter is settled.”*
Say:
*“this failure surface triggered this remedy family, these beneficiaries receive these components, this record explains the compensatory amount, this bundle carries the non-cash or collective repair, and these questions remain open after acceptance.”*
An offer is not repair.
Remedy needs structure. reacted to SoulInPsyAbstract's post with 🔥 about 4 hours ago
Tech digest, ~24h (Aug 19–20, 2026) — SIPA OS lens
One I actually checked myself, not took on faith: ShadowRay 2.0 is real. CVE-2025-62593, CVSS 9.4, browser-triggered RCE via DNS rebinding, actively hunting GPU clusters to turn into crypto-mining botnets. CISA gave federal agencies a 3-day patch window ending today.
Checked what "checked" actually covers, because that's the whole point of this series: no Ray package, process, port, or container on my always-on server — but the real training runs on ephemeral GPU instances that spin up per job and get torn down after. Those weren't up to inspect. So I went to the actual source instead: grepped all 10 training scripts that run on those instances for any import ray / ray.init / ray.train / ray.tune. Zero matches — the stack there is HuggingFace transformers/PEFT/bitsandbytes, nothing Ray-shaped in it. Code-level check beats an instance check anyway: if it's not imported, it can't run, whether or not the box is up.
Not affected, and now I can say why, not just that.
The rest is reported, not independently verified by me — treat accordingly:
* DeepSeek V4 Pro left preview, price jumped ~14x over V4 Flash. V4 Flash (open weights, end of July) is reportedly closing the gap on Opus 4.8 for code — if true, the flash tier gets more interesting than the pro tier for a lot of workloads.
* MCP moved under the Linux Foundation. If the reported 1M+ repos already importing an MCP SDK is accurate, this stopped being "an interesting agent protocol" and became infrastructure you build on, not around.
* Cursor pushed agents toward always-on: persistent goals, background subagents on isolated VMs instead of request/response. Matches where I'd bet agent tooling has to go — event-driven, not chat-driven.
* CISA also patched a Copilot memory-poisoning + one-click exfil bug (dubbed CoSnitch) — a reminder that "memory" as a feature is also an attack surface the moment it's writable by untrusted input. reacted to FredyRivera-dev's post with 🔥 about 4 hours ago
I've written a technical blog post about how we create a multimodal model:
Kairos: a multimodal model built with LFM2.5-2.6B as the LLM, MoonViT-3D (the vision tower of Kimi-K2.6) as the vision encoder, and a custom projector.
The original plan was LLaVA's approach, two stages: first align the projector with the LLM frozen, and then train the projector + LLM together. The first stage worked in terms of loss (ablation with +3.7 nats in favor of the image), but in free generation the image shifted the logits without changing the argmax: the model received the image and ignored it. That's why we jumped directly to early fusion, with a reasoning dataset.
For that, we created Kairos-Multimodal-Reasoning: 116,357 examples with explicit reasoning traces, generated through distillation (60,041 from LLaVA-CC3M, 2,295 from WebSight, and 54,021 from Zebra-CoT), with GPT 5.6 Luna, Inkling, Qwen 3.6 27B, and Qwen 3.7 Plus as teachers.
The training, in two phases:
1. Projector through backbone with 80k image-text pairs (Kairos-Proj-80k).
2. Projector + LoRA (r=16) with 30k examples from the reasoning dataset (Kairos-Alig-30k).
Everything is open source:
- Full blog post with the process: https://aquiles-ai.vercel.app/blog/kairos-a-multimodal-model
- Implementation: https://github.com/Aquiles-ai/Kairos
To be honest: the checkpoints are not a competent model, they are experimental artifacts. But they validated the approach and precisely defined what the next iteration needs.
https://huggingface.co/collections/Aquiles-ai/kairos
https://huggingface.co/Aquiles-ai/MoonViT-3D
https://huggingface.co/LiquidAI/LFM2.5-2.6BOrganizations
