<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Sugam Panthi</title><description>Writing on LLM evaluation, tool-calling agents, and how people interact with LLMs.</description><link>https://spanthi.com/</link><language>en-us</language><item><title>None of ARR’s 69,781 Published Paper Scores Is a 5.0</title><link>https://spanthi.com/blog/five-years-zero-fives/</link><guid isPermaLink="true">https://spanthi.com/blog/five-years-zero-fives/</guid><description>I pulled every score ACL Rolling Review has published: 69,781 scored submissions across 35 cycles, 2021 to 2026. No paper ever averaged a 5.0. The scores are per-paper aggregates, not individual reviews, the mode is 2.5, and the meta score swings wider than a paper&apos;s averaged reviews.</description><pubDate>Thu, 20 Aug 2026 00:00:00 GMT</pubDate><category>Peer Review</category><category>ACL Rolling Review</category><category>ARR</category><category>Score Distribution</category><category>Meta-Review</category><category>NLP</category><category>Data Analysis</category></item><item><title>Every Reported Number Should Trace to One Scorer and Run</title><link>https://spanthi.com/blog/a-number-you-cant-trace-is-a-rumor/</link><guid isPermaLink="true">https://spanthi.com/blog/a-number-you-cant-trace-is-a-rumor/</guid><description>The workflow I use across concurrent paper projects: each reported value points to a scorer and run, stop conditions precede spending, full runs require five checks, logs are append-only, and detector outputs receive a manual audit.</description><pubDate>Tue, 18 Aug 2026 00:00:00 GMT</pubDate><category>Research Workflow</category><category>Reproducibility</category><category>Provenance</category><category>Experiment Tracking</category><category>Research Engineering</category></item><item><title>LLMs Absorb Text Typed After a Paste</title><link>https://spanthi.com/blog/where-does-the-paste-end/</link><guid isPermaLink="true">https://spanthi.com/blog/where-does-the-paste-end/</guid><description>Across 19 models, bare paste seams absorb typed afterthoughts at rates from 8% to 68%. Whitespace does not help; provenance markup does, and matching the artifact’s voice makes the failure sharply worse.</description><pubDate>Sat, 15 Aug 2026 00:00:00 GMT</pubDate><category>LLMs</category><category>Prompting</category><category>Evaluation</category><category>SEAM</category><category>Human–AI Interaction</category></item><item><title>Recovery Tools Help Agents Respond to Corrupted Tool Results</title><link>https://spanthi.com/blog/agents-believe-tools-that-lie/</link><guid isPermaLink="true">https://spanthi.com/blog/agents-believe-tools-that-lie/</guid><description>Across five models, a nonbinding monitor improved recovery from corrupted tool results when its receipt named callable alternatives. Detection alone did not produce the measured gain.</description><pubDate>Sun, 09 Aug 2026 00:00:00 GMT</pubDate><category>Agents</category><category>Tool Use</category><category>LLMs</category><category>Outcome Monitors</category><category>Runtime Verification</category></item><item><title>Memory Consolidation Turns Temporary Statements Into Standing Facts</title><link>https://spanthi.com/blog/aspect-persistence-eternal-present-memory/</link><guid isPermaLink="true">https://spanthi.com/blog/aspect-persistence-eternal-present-memory/</guid><description>An initial 2,000-call LAPSE pilot found that consolidation selectively rewrote temporary forms into standing facts, while date stamps and a one-line preservation instruction barely changed later behavior.</description><pubDate>Mon, 03 Aug 2026 00:00:00 GMT</pubDate><category>LLM Memory</category><category>Temporal Reasoning</category><category>LAPSE</category><category>Memory Consolidation</category></item><item><title>Scripted product demos with GSAP, part two: the camera</title><link>https://spanthi.com/blog/gsap-choreography-part-2/</link><guid isPermaLink="true">https://spanthi.com/blog/gsap-choreography-part-2/</guid><description>Part one replaced a 3.4 MB video with 40 KB of DOM animation. Part two adds a camera: zoom into the action, pan to follow the cursor, pull back for scene changes. Plus the SEO, accessibility, and agent arguments I skipped the first time.</description><pubDate>Tue, 26 May 2026 00:00:00 GMT</pubDate><category>GSAP</category><category>Animation</category><category>React</category><category>Frontend</category><category>Performance</category><category>SEO</category><category>Accessibility</category></item><item><title>Scoring Targets Change Which Memory System Wins</title><link>https://spanthi.com/blog/your-memory-benchmark-is-lying-to-you/</link><guid isPermaLink="true">https://spanthi.com/blog/your-memory-benchmark-is-lying-to-you/</guid><description>A public companion to our arXiv paper on LLM memory evaluation: the same ranked list can produce different winners when Raw, Source, and Canonical scoring targets are swapped.</description><pubDate>Fri, 22 May 2026 00:00:00 GMT</pubDate><category>LLM Memory</category><category>Evaluation</category><category>Information Retrieval</category><category>Benchmarks</category><category>LongMemEval</category><category>LoCoMo</category><category>TIAP</category></item><item><title>Scripted product demos with GSAP instead of video</title><link>https://spanthi.com/blog/gsap-choreography/</link><guid isPermaLink="true">https://spanthi.com/blog/gsap-choreography/</guid><description>A 15-second hero video was 3.4 MB, letterboxed on phones, and ignored prefers-reduced-motion. I replaced it with a scripted GSAP animation: one timeline, cursor choreography, scene transitions, under 40 KB gzipped.</description><pubDate>Wed, 20 May 2026 00:00:00 GMT</pubDate><category>GSAP</category><category>Animation</category><category>React</category><category>Frontend</category><category>Performance</category></item><item><title>Voice Agents Need a Record of What the User Heard</title><link>https://spanthi.com/blog/every-voice-ai-forgets-what-you-heard/</link><guid isPermaLink="true">https://spanthi.com/blog/every-voice-ai-forgets-what-you-heard/</guid><description>A protocol proposal for voice-agent interruptions: generated audio races ahead of playback, current stacks expose cancellation and truncation, and agents still need first-class heard state and resumable threads.</description><pubDate>Sun, 03 May 2026 00:00:00 GMT</pubDate><category>Voice AI</category><category>Realtime</category><category>Agents</category><category>OpenAI</category><category>LiveKit</category><category>Pipecat</category><category>Protocol Design</category></item><item><title>Evidence Compression Is Reader-Dependent</title><link>https://spanthi.com/blog/compression-is-a-coin-flip/</link><guid isPermaLink="true">https://spanthi.com/blog/compression-is-a-coin-flip/</guid><description>In the SIEVE study, the same compressed evidence helped weaker reader models more than stronger ones. For the strongest reader, the help-to-damage ratio fell to 1:1 and 15% of previously correct answers became wrong.</description><pubDate>Fri, 24 Apr 2026 00:00:00 GMT</pubDate><category>LLM Memory</category><category>RAG</category><category>Evidence Compression</category><category>Evaluation</category><category>SIEVE</category><category>LongMemEval</category><category>HotpotQA</category></item><item><title>Technical SEO for a Multilingual Browser Game</title><link>https://spanthi.com/blog/seo-optimization/</link><guid isPermaLink="true">https://spanthi.com/blog/seo-optimization/</guid><description>An implementation report for playimposter.xyz covering page-specific metadata, canonical URLs, multilingual hreflang clusters, structured data, crawl directives, sitemaps, and server configuration.</description><pubDate>Tue, 03 Mar 2026 00:00:00 GMT</pubDate><category>SEO</category><category>Technical SEO</category><category>Schema</category><category>Hreflang</category><category>Web Performance</category></item><item><title>Building deepseek-go, Part 1: Design Constraints</title><link>https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-1/</link><guid isPermaLink="true">https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-1/</guid><description>Part 1 of my deepseek-go series: why I built it, the production problems it solved, and the architecture principles that shaped the package.</description><pubDate>Tue, 03 Mar 2026 00:00:00 GMT</pubDate><category>Go</category><category>SDK</category><category>API</category><category>DeepSeek</category><category>Open Source</category></item><item><title>Building deepseek-go, Part 2: Architecture and API Design</title><link>https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-2/</link><guid isPermaLink="true">https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-2/</guid><description>Part 2 dives into deepseek-go internals: client options, request/response handlers, streaming, FIM, JSON mode, and provider compatibility.</description><pubDate>Tue, 03 Mar 2026 00:00:00 GMT</pubDate><category>Go</category><category>SDK</category><category>Architecture</category><category>Streaming</category><category>API Design</category></item><item><title>Building deepseek-go, Part 3: How Adoption Changed Maintenance</title><link>https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-3/</link><guid isPermaLink="true">https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-3/</guid><description>Part 3 analyzes adoption and growth with live GitHub stats, release cadence, star-history graphs, and what those trends changed in my engineering priorities.</description><pubDate>Tue, 03 Mar 2026 00:00:00 GMT</pubDate><category>Go</category><category>Open Source</category><category>GitHub</category><category>Metrics</category><category>Engineering Leadership</category></item><item><title>Building deepseek-go, Part 4: Maintainer Lessons and Roadmap</title><link>https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-4/</link><guid isPermaLink="true">https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-4/</guid><description>Part 4 closes the series with contributor thanks, hard SDK lessons, and my roadmap for deepseek-go&apos;s next stage.</description><pubDate>Tue, 03 Mar 2026 00:00:00 GMT</pubDate><category>Go</category><category>Open Source</category><category>Maintainer</category><category>Roadmap</category><category>Community</category></item><item><title>Building deepseek-go, Part 5: Testing the Client</title><link>https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-5-testing/</link><guid isPermaLink="true">https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-5-testing/</guid><description>Part 5 details my Go testing strategy for deepseek-go: layered tests, stream safety checks, and release confidence practices.</description><pubDate>Tue, 03 Mar 2026 00:00:00 GMT</pubDate><category>Go</category><category>Testing</category><category>SDK</category><category>Quality</category><category>Open Source</category></item><item><title>Building deepseek-go, Part 6: The Ollama Adapter</title><link>https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-6-ollama/</link><guid isPermaLink="true">https://spanthi.com/blog/creating-most-popular-deepseek-api-client-go-part-6-ollama/</guid><description>Part 6 explains how I integrated Ollama into deepseek-go with explicit adapters, stream resilience, and practical provider compatibility.</description><pubDate>Tue, 03 Mar 2026 00:00:00 GMT</pubDate><category>Go</category><category>Ollama</category><category>SDK</category><category>LLM</category><category>Integration</category></item><item><title>What Edge Deployment Changed About My ML Work</title><link>https://spanthi.com/blog/reflections-on-ai-internship/</link><guid isPermaLink="true">https://spanthi.com/blog/reflections-on-ai-internship/</guid><description>At Prediction3d, edge deployment made data quality, memory use, latency, quantization, pruning, and exported-runtime checks part of the model-selection decision.</description><pubDate>Tue, 10 Dec 2024 00:00:00 GMT</pubDate><category>AI</category><category>Machine Learning</category><category>Career</category></item><item><title>Getting Started with Go for Web Services (Part 3): Testing, Deployment, Scaling</title><link>https://spanthi.com/blog/getting-started-with-go-testing-deploy-scaling/</link><guid isPermaLink="true">https://spanthi.com/blog/getting-started-with-go-testing-deploy-scaling/</guid><description>Part 3 of the Go web services series: testing pyramid, deployment safety checks, rollout strategy, scaling sequence, and production checklist.</description><pubDate>Thu, 07 Nov 2024 00:00:00 GMT</pubDate><category>Go</category><category>Testing</category><category>Deployment</category><category>Scaling</category><category>SRE</category></item><item><title>Getting Started with Go for Web Services (Part 2): Data, Middleware, Auth</title><link>https://spanthi.com/blog/getting-started-with-go-data-middleware-auth/</link><guid isPermaLink="true">https://spanthi.com/blog/getting-started-with-go-data-middleware-auth/</guid><description>Part 2 of the Go web services series: repository boundaries, middleware ordering, authentication, database handling, and observability.</description><pubDate>Wed, 06 Nov 2024 00:00:00 GMT</pubDate><category>Go</category><category>API</category><category>Middleware</category><category>Auth</category><category>Observability</category></item><item><title>Getting Started with Go for Web Services (Part 1): Foundations</title><link>https://spanthi.com/blog/getting-started-with-go/</link><guid isPermaLink="true">https://spanthi.com/blog/getting-started-with-go/</guid><description>Part 1 of a detailed Go web services series: setup, server lifecycle, configuration, error handling, and Python-to-Go workflow.</description><pubDate>Tue, 05 Nov 2024 00:00:00 GMT</pubDate><category>Go</category><category>Backend</category><category>API</category><category>Web Services</category><category>System Design</category></item><item><title>Building My Portfolio Aesthetic</title><link>https://spanthi.com/blog/building-my-portfolio-aesthetic/</link><guid isPermaLink="true">https://spanthi.com/blog/building-my-portfolio-aesthetic/</guid><description>How I designed a paper-aesthetic texture and integrated it into a modern React portfolio using Tailwind CSS.</description><pubDate>Fri, 25 Oct 2024 00:00:00 GMT</pubDate><category>React</category><category>Tailwind</category><category>Design</category></item></channel></rss>