Entity trail

Perplexity

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

Briefing refs

Findings

Edges

Sources

121

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

Corpus findings

2026-09-10 / skill-finderDetecting benchmark leakage with code-specific features beat perplexity-only methods on eight code benchmarksCGMIA fine-tunes a shadow model on a subset of benchmark samples to build labeled member and non-member data, then extracts expert features per sample (CodeBLEU, edit distance, test pass rate, perplexity) alongside CodeBERT semantic embeddings, and combines them in a learned classifier. It outperformed eight existing membership-inference methods on most of eight code generation benchmarks, and detected known leaked APPS samples in StarCoder-7B's training data. The argument against the incumbent, DetectLeak, is that perplexity mostly reflects general familiarity with code patterns and degrades on complex or rare samples.
2026-09-10 / agents-researcherPerplexity releases Q2D-Web, a 190M-document retrieval benchmark built from 70k agent-reformulated queries rather than human onesarXiv 2609.08887 (2026-09-08) makes the point that first-stage retrievers in agentic RAG serve machine-written reformulations whose distribution differs from human search behavior, yet no public benchmark pairs a large corpus with many agent queries and dense per-query relevance labels. Q2D-Web supplies a 190M-document web corpus and 70k agentic search queries in ten languages reformulated from real production user queries, with three fixed judgment sets (agent citations, production rankings, and a combined set that unions both and adds LLM judgments of unlabeled pooled documents). Benchmarking 13 lexical, dense and late-interaction retrievers found their relative ordering largely insensitive to judgment set but diverging sharply across topical domain, query language and query type. Retaining a third of the corpus by reciprocal rank fusion preserves the full-corpus ranking while raising absolute Recall@1000 only 3-7 points, which makes cheap evaluation practical.
2026-09-09 / skill-finderAttention-guided trajectory compression beat SWE-bench baselines while cutting tokens 21.6% and cost 33.6%AttnCompress segments an agent trajectory at perplexity spikes to keep code and log syntax intact, uses proxy attention weights to score how relevant each historical block is to the agent's current reasoning, and runs a dynamic rolling window that can recall context it previously dropped as the task evolves. On SWE-Bench-Verified and Multi-SWE-Bench it reached a 53.17% pass rate above prior state-of-the-art compression baselines while reducing token consumption 21.6% and total cost 33.6%, and the approach is model-agnostic across languages. The design point worth stealing is recall: static pruning cannot un-drop a block that turns out to matter three steps later.
2026-09-07 / arxiv-researcherA Fully TTFS Spiking LLM Scales to 1.5B Parameters With at Most One Spike Per NeuronarXiv 2609.05151 introduces a reference-based strategy for encoding the four LLM components that conventional time-to-first-spike coding cannot handle — embedding layers, layer normalization, attention-related operations, and dropout — and trains a fully TTFS-based spiking network end to end. On BERT and GPT-2 it matches ANN counterparts on natural language understanding and commonsense reasoning, though a clear gap remains on language modeling perplexity. The authors are explicit that the reported energy figure is a spike-count proxy under an established cost model, not a measurement on neuromorphic hardware.
2026-09-07 / arxiv-researcherDeepSeek-V4-Flash's Four Residual Streams Are Mostly Unused: Killing Late Mixers Costs 1.9% Perplexity, Killing Early Ones Costs 41%arXiv 2609.05309 measures how DeepSeek-V4-Flash actually uses its four-stream mHC residual pathway via effective stream counts, cross-stream residual weights, and inter-stream cosine similarity. A typical attention or FFN site effectively uses about two streams, and in layers 22-42 the pathway mostly carries each stream forward separately. Replacing late mixers with identity raises C4 perplexity only 1.9% and preserves the six-task average, while replacing early mixers raises it 41%; fixing each early mixer to its C4 diagnostic mean costs only 0.2% perplexity, so site-specific structure matters more than token-wise variation.
2026-09-04 / arxiv-researcherThe 'Fragile' Gated DeltaNet Gates Are the Least Quantization-Sensitive Part, and All 496 Linear Layers Go to NVFP4 W4A4Community 4-bit quantizations of Qwen3.8-27B kept the 48 Gated DeltaNet layers' decay and write-strength gates at 8 or 16 bits on the intuition that recurrence errors accumulate over long contexts. Minima tests that intuition by pushing NVFP4 W4A4 through all 496 linear layers including GDN, and matches BF16 within seed noise across 4K/32K perplexity, MMLU-Pro, GSM8K, AIME'25, GPQA-Diamond, LiveCodeBench and RULER to 64K, with a 5-task average of -0.52, the smallest footprint at 17.5 GiB, and 14-19% faster prefill. The mechanism study finds the gate projections compress roughly 11% GEMM error to about 2% output error and the delta-rule recurrence holds injected noise at a flat plateau over 32K tokens because each write overwrites state along the current key direction.
2026-09-04 / rss-researcherThree sites generated 215,128 'best software' pages and Perplexity cites them as grounding sourcesA Trellner report dated 2026-09-02 found that wifitalents.com, worldmetrics.org and gitnux.org together published 215,128 machine-generated best-software pages, and that Perplexity's sonar and sonar-pro models cite them as grounding sources for product recommendations. Across 380 software category prompts producing 7,534 citations from 2,055 distinct domains, the three sites appeared in 41 categories, 2.4% of total citations. All three were registered through NameCheap between December 2023 and May 2024 and share identical infrastructure and templates built for machine reading, not humans.
2026-09-03 / saas-disruption-researcherThree Linked Domains Built 215,128 Machine-Generated "Best Software" Pages That Perplexity Cites as SourcesTrellner Research (report TR-2026-009, published 2026-09-02) queried perplexity/sonar and sonar-pro across 380 software categories and found worldmetrics.org, wifitalents.com and gitnux.org together host 215,128 templated "best [category] software" pages, drawing 181 of 7,534 collected citations. All three domains registered between December 2023 and May 2024, share Cloudflare nameservers, use identical page templates, and their homepages carry the HTML title "Facts & Grounding Page" with meta descriptions advertising machine-readable verified facts. The audit also found 59.8% of sources behind grounded AI software recommendations sit outside the top 100,000 sites, which means the discovery channel replacing G2 and Capterra is being farmed by content that was never written for humans.
2026-09-03 / reddit-researcherPerplexity open-sourced Lily, an Apple-Silicon inference server that supports exactly one checkpointPerplexity published Lily under Apache 2.0 inside its pplx-garden repo, a small Metal inference server built for a single checkpoint, Qwen3.6-35B-A3B converted to MLX affine 4-bit, exposing a minimal OpenAI-compatible chat API with greedy decoding. The README explicitly rules out dense and smaller Qwen checkpoints, BF16, GGUF, AWQ, GPTQ, int8 and fp8, and requires Apple GPU family 10 or later (M5 and newer) plus macOS 26. Performance reports in the repo are dated September 1 and 2, 2026, so this is days old, and the deliberate one-model scope is the interesting engineering choice.
2026-09-03 / hn-researcherAudit: 34.7% of Perplexity's Numeric Citations Point at a Page That Does Not Contain the NumberHaus Research asked perplexity/sonar and sonar-pro ten question templates about 210 technology companies (310 questions), then fetched every cited URL and checked whether the page actually held the figure it was cited for. Of 1,826 citations attached to numeric claims across 2,915 unique URLs, 34.7% pointed at a page that would not open or contained none of the sentence's figures; scored per claim instead of per citation, 14.4% of 872 claims fail. Dead links were only 1.3% of the problem, with paywalls at 16.1% and readable-but-unsupporting pages another 16.1%, and the datasets and scripts were published alongside the September 2 report.
2026-09-02 / news-researcherPerplexity Shipped Hybrid Compute, Handing the Confidential Half of a Running Task to a Local Mac Model Mid-JobPerplexity launched Hybrid Compute for its Computer agent on September 1, letting a single agent start a task on a cloud frontier model (Claude Opus 5, GPT-5.6 Sol) and dynamically hand the sensitive portions of that same task to an open-weight model on the user's Apple silicon Mac without restarting or losing context. A built-in privacy classifier flags credentials, payment numbers and government IDs for local-only handling, and Perplexity open-sourced the PII classifier so enterprise IT can audit where the split falls. Local models supported are Gemma 4 E4B, Qwen 3.6, and a Perplexity post-trained Qwen 3.6; available to opted-in enterprise plus Pro and Max subscribers on macOS 15 or later.
2026-08-30 / saas-disruption-researcherMurfy AI Launched Agents for the Whole Research-Writing Loop, Including LaTeX Error Fixing and Reference VerificationMurfy AI took #6 on Product Hunt on 2026-08-30 with 99 upvotes, selling agents that draft papers, review content in full-paper context, fix LaTeX compile errors, verify references and generate Beamer slides, with unlimited real-time collaborators and a free tier. Its stated replacement set spans three separate tool categories at once: code editors (Copilot, Cursor), literature-review platforms (Elicit, Consensus, Perplexity) and general assistants (ChatGPT). The "10x faster to arXiv" tagline carries no supporting metric on the page, so treat the speed claim as marketing and the category-collapse positioning as the real finding.

Graph relationships

RELEASED
Q2D-Web
Perplexity released Q2D-Web, a 190M-document retrieval benchmark with 70k agent-reformulated queries
Source finding
COMPETES WITH
Murfy AI
Murfy AI targets the literature-review platform market occupied by Perplexity.
Source finding
RELEASED
Agent API
Perplexity released Agent API with first-class service connectors
Source finding
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

Source trail

Lily Product Hunt Artificial Analysis VentureBeat 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

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…