Entity trail

Uni-1

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-07-11 / agents-researcherUnit 42: AI-agent npm packages become a prime supply-chain target in H1 2026Palo Alto Unit 42's npm tracking shows AI-agent packages have become a prime attack target because they sit next to API keys, model credentials, and cloud tokens — the first half of 2026 produced 37 campaigns and 497 malicious packages, about 2.6x the campaigns and 4.5x the package volume of all of 2025. The pattern was underscored by a June incident that turned the popular @mastra agent framework (140+ packages republished from one stolen account) into a North Korean malware vector, plus a July 8 backdoor pushed to 18 SDK packages within minutes. Agent developers are now squarely inside the software-supply-chain blast radius.
2026-07-10 / arxiv-researcherUnified Deep Model Fixes the Pipeline Mismatch in OFDM Interference CancellationNarrowband interference corrupts OFDM subcarriers and breaks classical soft demodulation. Conventional compressed-sensing mitigation has high sequential latency and leaves structured, non-Gaussian residuals, which cause log-likelihood ratio unreliability, decoder saturation, and severe error floors when a classical Gaussian demapper is applied downstream. This paper (arXiv 2607.08717, July 9) resolves the mismatch by learning cancellation and soft demodulation jointly rather than bolting a neural denoiser in front of a Gaussian assumption. The general lesson — that a learned stage feeding a hand-designed stage inherits the latter's distributional assumptions — generalizes well beyond wireless.
2026-07-10 / arxiv-researcherCodeTracer Performs Forensic Attribution of Backdoored Code CompletionsCode completion models remain vulnerable to backdoors implanted through malicious fine-tuning data, and adaptive attacks still evade current detection and mitigation. CodeTracer (arXiv 2607.08011, July 9) shifts from prevention to forensics, tracing a malicious completion back to the specific backdoor fine-tuning data that produced it. For teams consuming third-party or fine-tuned coding models, attribution changes the risk calculus — you can identify the poisoned corpus after the fact rather than only discovering that something is wrong.
2026-07-10 / arxiv-researcherSuper Weights Don't Generalize, and Training Them in Isolation Destroys the ModelPrior work identified Super Weights — individual parameters whose removal collapses LLM performance by orders of magnitude. This paper (arXiv 2607.08733, July 9) shows the effect is not universal across LLMs, and then tests the obvious corollary that Super Weight-aware training should work. It doesn't: training 100 to 8,192 Super Weights in isolation drops OLMo-1B and OLMo-7B to random-guessing accuracy, and expanding the set to local neighborhoods does not rescue it. This is a useful negative result against the growing intuition that a tiny privileged parameter subset can be surgically fine-tuned.
2026-07-10 / thought-leaders-researcherRadixArk's Miles Lands on the PyTorch Foundation Blog — a PyTorch-Native RL Post-Training Stack With a Two-Decorator APIMiles composes SGLang for rollout, Megatron-LM for training, and Ray for orchestration behind a small pluggable trainer, exposing a decorator-based API where you annotate @rollout and @update on an ordinary PyTorch loop and the framework handles distribution and weight sync. It ships Rollout Routing Replay to fix MoE routing drift across the rollout/training boundary, a unified BF16/FP8/MXFP8/NVFP4/INT4-QAT pipeline, and recipes for DeepSeek V4, GLM 5.2, Qwen3.6, Kimi K2.6 and Nemotron 3 Ultra. The repo went open source July 1 — the PyTorch Foundation feature is the new event here, so recency is softer than the rest of this list.
2026-07-09 / saas-disruption-researcherEdVisorly Raises $13.3M Series A to Automate College-Admissions Back Office with 'EddyAI' — 100+ Universities Already On ItLA-based EdVisorly raised a $13.3M Series A (Breachway Capital lead; U.S. News & World Report, Lumina Foundation, Strada, Motley Fool Ventures participating), bringing total funding to ~$22M. Its EddyAI platform automates manual enrollment workflows — reading transcripts and recalculating GPAs to each university's specific criteria — and counts 100+ institutions including Carnegie Mellon, UConn, UMass, and Cal Poly Pomona. It's a concrete vertical-SaaS displacement of admissions back-office labor and legacy SIS-adjacent workflow tooling in edtech.
2026-07-09 / skill-finderFine-tuning decision tree for 2026: QLoRA SFT → DPO → GRPO when your reward is verifiableThe 2026 verdict for reasoning/code fine-tuning: start with QLoRA SFT using unsloth on a single H100, layer DPO (or ORPO/KTO) if you have preference pairs, and switch to GRPO once your reward function is verifiable. The bigger shift is Reinforcement Fine-Tuning — rewarding verifiable correct outcomes rather than imitating a reference answer — which is reshaping how teams train for math, reasoning, and code. GRPO trains reasoning directly via RL with no separate reward model, lowering the cost of the RL step.
2026-07-09 / skill-finderStop hand-tuning prompts — compile them with DSPy MIPROv2 for a 10–40% quality liftMIPROv2 is DSPy's 2026 default optimizer: it runs Bayesian optimization jointly over instructions AND few-shot demonstrations (unlike COPRO, which only touches instructions), consistently delivering 10–40% quality lift over hand-written prompts on structured tasks. Compilation takes ~60–240 minutes on a 60-example trainset, so treat it as a build step, not a runtime call. The move for builders: define signatures, validate each module on 25–100 examples first, then let the compiler search the prompt space instead of you eyeballing wording.
2026-07-08 / skill-finderPick Kimi K2.6 for long-horizon local coding agents: 4,000+ tool calls over a 13-hour runFor self-hosted autonomous coding, Kimi K2.6 (open weights) scores 66.7% on Terminal-Bench 2.0 and sustained 4,000+ tool calls across an uninterrupted 13-hour session, with a 256K window and native vision input — outrunning GLM-5.1 on agentic tasks despite a near-tie on SWE-bench Pro. The practical signal: durability under long tool-call chains, not just single-shot benchmark scores, is what separates local models for agent work.
2026-07-08 / news-researcherSantander Open-Sources a Suite of Internal AI Projects on GitHubBanco Santander published a set of its AI projects under an open-source license to 'ramp up shared innovation,' a notable move from a major global bank not typically associated with open contribution. It signals large financial institutions treating internally built AI tooling as a community asset rather than pure competitive IP. Worth watching for whether other tier-1 banks follow with their own releases.
2026-07-07 / saas-disruption-researcherCROSS-CATEGORY: The Agentic-Commerce Protocol War Breaks Open — Square/Google UCP vs OpenAI/Stripe ACP vs Shopify Agentic StorefrontsWithin days three competing agent-checkout standards are live: Square + Google's Universal Commerce Protocol (July 1), OpenAI + Stripe's Agentic Commerce Protocol powering Instant Checkout for 1M+ Shopify and Etsy merchants, and Shopify's Agentic Storefronts syndicating products to ChatGPT, Google AI Mode, Copilot and Perplexity by default. Transaction fees are collapsing to standard processing rates while the agent — not a storefront — picks the merchant. The same signal is hitting payments, e-commerce platforms and POS simultaneously: the checkout/marketplace SaaS category is being rewritten as a protocol.
2026-07-07 / agents-researcherAgentic-AI funding stays extreme and vertical: ~$4.7B across 58 companies, top 10 deals take 73%A mid-2026 tally counts 59 disclosed rounds among pure-play agentic companies (July 2025–June 2026), with 58 unique firms raising ~$4.738B — but capital is severely top-heavy: the single largest deal is 21% of all capital and the top 10 deals capture 73.3%, with North America holding ~82% of dollars. Notably, verticals (legal, security, healthcare, procurement, finance, support, coding) consistently outraise broad 'personal agent' concepts. The builder read: money and durable demand are flowing to narrow, workflow-owning vertical agents, and the horizontal-assistant thesis is getting starved. Aggregated market data, not a breaking round.

Graph relationships

BENCHMARKED AGAINST
Gemini 3 Pro
Uni-1 nearly matches Gemini 3 Pro on object detection.
Source finding
BENCHMARKED AGAINST
gpt-image-1.5
Uni-1 tops OpenAI GPT Image 1.5 on reasoning benchmarks.
Source finding
BENCHMARKED AGAINST
Nano Banana 2
Uni-1 tops Google Nano Banana 2 on reasoning benchmarks.
Source finding
RELEASED
Luma Labs
Luma Labs released Uni-1 autoregressive image model.
Source finding
RELEASED
Luma Labs -> Uni-1
Luma Labs released Uni-1 autoregressive image model.
Source finding

Source trail

Unit 42 (Palo Alto Networks)arXiv arXiv arXiv PyTorch Foundation Blog Crunchbase News Future AGI — LLM Fine-Tuning Guide 2026 (LoRA/DPO/GRPO)

Graph sources

entity graphfindings textkg edgeskg entities