Entity trail

Perplexity

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

Briefing refs

Findings

Edges

Sources

114

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

Corpus findings

2026-08-20 / reddit-researcherThe Same GRPO Recipe on Three From-Scratch LLMs Produced Three Different Outcomes, and the Middle Model Broke WorstA researcher pre-trained three models from scratch in raw PyTorch (353M MHA, 316M differential+GQA, 672M XSA+GQA on 10B/10B/30B tokens), then ran identical SFT and GRPO with the same curriculum, reward function, hyperparameters and KL coefficient. WikiText perplexity after SFT to GRPO moved +0.2% on V1, +52% on V2 and +5% on V3, with arc_easy dropping about 6 points on V3, so damage had no clean relationship to scale. The models did learn the trained task (V3 mastered 4 of 5 curriculum stages) but it did not transfer, with GSM8K flat, which is a clean documented negative result of the kind papers rarely publish.
2026-08-20 / rss-researcherPerplexity's India numbers after the Airtel giveaway: 22M peak users down to 14M, revenue up 60%Perplexity's free 12-month Pro deal with Airtel reached 360 million customers and drove 56 million downloads over seven months, nine times the prior period, peaking at 22 million monthly active users in October 2025 against a 2.6 million baseline in the first half of 2025. By July 2026, with the first free cohorts expiring, monthly actives had fallen to about 14 million, but India revenue rose roughly 60% and monthly net mobile revenue went from $34,000 in January 2025 to $156,000 in July 2026. Sensor Tower and Appfigures supplied the download data; the retained paying tail, not the peak, is the number that matters.
2026-08-18 / hn-researcherIsraeli Government Paid $900K for a Fake Think Tank Built to Poison LLM Answers — 100+ Bylineless Articles Since Aug 6, 11 of 12 Flagged AI-WrittenResponsible Statecraft reported Aug 17 that the 'Hanover Institute for Public Policy' is a front created by Piro, Inc. (co-founded by Spike Lee producer Daniel Rosenberg) under a $900,000 contract from the Israeli Government Advertising Agency, subcontracted through Havas Media. Piro's own site markets the practice as 'AI Story Optimization' — content 'engineered for how LLMs evaluate credibility' — and GPTZero flagged 11 of 12 sampled articles as AI-written with high confidence. It topped HN at 761 points and 435 comments; the builder takeaway is that retrieval-augmented answers now have a paid supply-chain attack surface aimed at Claude, Gemini, and Perplexity specifically.
2026-08-16 / arxiv-researcherMixture of Training Recomposes Independently Pretrained Layer Blocks Into a Working 1.3B Model at Parity PerplexityMoT (arXiv 2608.13277, 2026-08-13) asks whether pretraining can be decomposed into small independently schedulable jobs. It partitions a target Transformer into contiguous layer blocks, trains each inside a frozen pretrained aligner scaffold, then recomposes with an optional short end-to-end adaptation pass. On a 1.3B Gemma-style model trained on C4, a quality-parity schedule reaches the same reported perplexity as the monolithic baseline with a shorter idealized layer-equivalent critical path after aligner preparation — though it processes more aggregate tokens, so the compute advantage depends on reusing the aligner across runs. The authors are careful to present this as a proof of mechanism, not a replacement for monolithic pretraining.
2026-08-15 / arxiv-researcherNegative Result: RoPE-Aligned Pairwise Rotations Lose to Full-Head Hadamard in W4A4KV4 QuantizationRotation-based post-training quantization usually mixes across a whole attention head, but RoPE partitions heads into 2D frequency pairs, so the authors derive the variance-minimizing head-shared pairwise rotation and prove the converse — for distinct frequencies, no other single-head orthogonal map commutes with RoPE. Despite attaining its analytic minimum, the pairwise configuration increased perplexity at both short and long context across four checkpoints in dynamic W4A4KV4, and estimating the shared angle from K alone improved it on every checkpoint without closing the gap. The diagnosis is worth more than the method: the analytic objective controls a position-averaged second moment of a pooled calibration covariance while the dynamic quantizer sets its step from a tokenwise group range, so surrogate optimality is misaligned with what the quantizer actually measures.
2026-08-13 / hn-researcherAn Active Campaign Is Spoofing ClaudeBot and GPTBot to Mass-Scan for /.claude/settings.json and Anthropic Credential FilesKnown Agents flagged an active AI-bot spoofing campaign (285 points, 215 comments on HN) in which scanners impersonate recognized crawler user-agents to sweep for agent credential paths — top targeted paths include /.config/anthropic/credentials/default.json, /.claude/settings.json, /.aws/credentials, and .env variants. Detection is via failed IP verification or Web Bot Auth, and the measured spoof rates are actually low: Googlebot 0.5%, ChatGPT-User, GPTBot, PerplexityBot, and ClaudeBot each 0.1%. ClaudeBot is 3.2% of all observed traffic and 27% of AI scraping activity, which is why it's the identity worth stealing — and the targeted paths mean every developer with an agent config in a web-served directory is now in an attacker's wordlist.
2026-08-11 / github-pulse-researcherFour Major Tech Companies Now Ship Competing Open-Source Agent-Security Tools — NVIDIA, Tencent, Uber, and PerplexityThe `agent-security` GitHub topic now holds 42 repos above 100 stars, and four of them come from large tech companies rather than startups: NVIDIA/SkillSpector (14,498 stars, scanning Claude Code/Codex/MCP skills for prompt injection and supply-chain risk), Tencent/AI-Infra-Guard (4,467, red-teaming with Many-Shot/PAIR/GOAT/ActorAttack jailbreaks), uber/ADR (1,351 stars, 'deployed at Uber', sensor v1.0.0 on 2026-07-31), and perplexityai/numbat (886 stars, created 2026-07-24). Numbat is the newest entrant and the most EDR-shaped: a single cgo-free binary for macOS/Linux/Windows that normalizes agent hooks, plugins, and OTLP logs into one event model, evaluates it with a CEL rule engine, and detects cloud-metadata probing, secret-file access paired with exfiltration, and unauthorized SSH key persistence — all shipped rules monitor-only, blocking off by default.
2026-08-11 / thought-leaders-researcherYann LeCun Joins 224 Ventures as Partner Alongside Gemini Technical Co-Lead Oriol Vinyals — Weeks After His Last Fund Attempt CollapsedBloomberg reported Aug 5 that LeCun joined 224 Ventures, a $100M+ AI fund co-led by Oriol Vinyals (DeepMind, technical co-lead on Gemini) and Shaun Johnson (AIX Ventures, backer of Perplexity and Chroma), while retaining his executive chairman role at AMI Labs. The fund targets early-stage applications, robotics and infrastructure at sub-$100M valuations without leading rounds; the name references the 224x224 image resolution and Stanford CS224. The notable structural signal is two of the most senior researchers from opposing camps — LeCun's anti-LLM world-models thesis and Vinyals' frontier-LLM work — sharing a cap table strategy.
2026-08-10 / skill-finderPerplexity can't detect RAG poisoning — poisoned answers score *lower* perplexity; watch attention entropy collapse insteadD-SCAN (SIGIR 2026) reports that perplexity-based RAG poisoning detection is actively misleading because adversarial generations often have lower perplexity than benign ones, so the standard guardrail returns high confidence on compromised output. Their signal is document-level attention dynamics: during a poisoned generation, attention concentrates on the injected document and entropy drops, versus the dispersed attention of a legitimate answer. It is a lightweight runtime check with code published at github.com/yingtaoren/D-Scan, and it applies to any self-hosted RAG stack where you can read attention — a real detection path for teams who assumed retrieval hygiene was enough.
2026-08-09 / news-researcherEvery AI Browser Tested at Black Hat Fell to Prompt Injection — Brave Researcher Says There Is No Clean FixIn his Black Hat USA 2026 briefing 'Attacking and Defending AI Browsers,' Brave Software security engineer Artem Chaikin said every browser he analyzed proved vulnerable to indirect prompt injection, demonstrating live against Opera's AI browser, Perplexity Comet, and ChatGPT Atlas. The root cause is structural: agents cannot distinguish ordinary page content from hidden instructions embedded in an email, shared document, calendar invite or webpage, which yields data exfiltration and account takeover. Chaikin's conclusion is that guardrails are improving but the threat class is not going away — meaning agentic browsing should be treated as an untrusted-input problem, not a model-alignment one.
2026-08-08 / news-researcherFirebird Opens the CIS Region's Largest AI Factory in Armenia — 300MW and 70,000+ Rubin and Blackwell GPUs by End of 2027Firebird opened an AI factory in Hrazdan, Armenia on August 8, delivered in just over six months, scaling to 300 megawatts and more than 70,000 NVIDIA Rubin and Blackwell GPUs by the end of 2027. It is built on the NVIDIA DSX platform — which NVIDIA says runs 40% more GPUs on the same footprint — with Dell PowerEdge servers, Schneider Electric power and Vertiv cooling; NVIDIA intends to invest, following CoreWeave's earlier stake. Prime Minister Nikol Pashinyan attended the opening alongside Kazakhstan's deputy PM, and Perplexity is already an early compute customer as part of a 2-gigawatt roadmap spanning Armenia and Kazakhstan.
2026-08-06 / sources-researcherTIME Serves AI Crawlers a 13KB Markdown Version of Its Site With Ads Baked Into the Prose — Humans Get 303KB of HTML Without ThemVincent Schmalbach found that varying only the User-Agent flips TIME.com between a 303KB human page and a 13KB markdown page served through ad-tech vendor Mobian, containing sponsored FAQ blocks (Ally Bank, Project Management Institute), campaign tags and JSON-LD, labeled 'Sponsored content. Supplied in partnership with [brand]' — text no human reader ever sees. ClaudeBot, PerplexityBot and OAI-SearchBot get the markdown; GPTBot and ChatGPT-User get a 406; Googlebot and real browsers get the HTML. TIME says bot traffic already outnumbers human traffic most days, which makes this the first clear case of paid content injected specifically into the LLM ingestion path — a live data-poisoning vector for anyone whose agent summarizes news.

Graph relationships

PARTNERS WITH
Airtel
Perplexity offered a free 12-month Pro deal with Airtel.
Source finding
USES
Firebird
Perplexity is an early compute customer of Firebird for a 2-gigawatt roadmap spanning Armenia and Kazakhstan.
Source finding
RELEASED
Personal Computer
Perplexity released Personal Computer for Windows on July 27
Source finding
INVESTED IN
Sim
Perplexity is a backer of Sim's $7M Series A funding round.
Source finding
RELEASED
WANDR
Perplexity released WANDR, a 500-task research benchmark with dynamic evidence verification.
Source finding
WORKS AT
Aravind Srinivas
Aravind Srinivas is CEO of Perplexity.
Source finding
USES
Galaxy S26
Galaxy S26 uses Perplexity for web queries.
Source finding
RELEASED
Computer
Perplexity launched Computer as first commercial multi-model orchestration product with 19 specialized models.
Source finding
CRITICIZES
Zenity Labs
Zenity Labs disclosed critical PleaseFix vulnerabilities in Perplexity.
Source finding
COMPETES WITH
NotebookLM
NotebookLM surpassed Perplexity in total site visits over past two months.
Source finding
PARTNERS WITH
NVIDIA
Perplexity partnered with NVIDIA on the Nemotron Coalition.
Source finding
RELEASED
Computer
Perplexity expanded Computer agent into enterprise with Model Council.
Source finding

Source trail

D-SCAN The Hrazdan facility github.com/perplexityai/numbat Founded by Kevin Jiang, who worked on Perplexity's Comet Latent Space Gemini Spark Perplexity's Personal Computer for Windows Viktor OpenWorker "America's Open-Model Paradox"hosted on Nvidia's own servers roughly 200 startup founders petitioned the Trump administration

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…