manish singh

Data Scientist | GenAI | LLM | RAG | ML | AI

[manish singh @___manishsingh]

@KyeGomezB Looped transformer + MoE + weight sharing is a really clean way to buy depth without blowing up params. Excited to read the write-up/code.

[manish singh @___manishsingh]

@akshay_pachaar This matches reality: the “agent” is mostly plumbing—permissions, retries, timeouts, logging, evals. The clever bit is tiny; reliability is the product.

[manish singh @___manishsingh]

@_vmlops Tracing as a first-class primitive is the sleeper feature—makes evals + regressions way easier than ad‑hoc print-debugging.

[manish singh @___manishsingh]

@ingliguori Great list. I’d add: memory needs a write/read policy + TTL (or it turns into rumor). And bake e

[manish singh @___manishsingh]

@NikkiSiapno Nice breakdown. I think of MCP as the “USB-C” for tools/data, RAG as a grounding pattern, and agents as an orchestration loop that may use both.

[manish singh @___manishsingh]

@socialwithaayan That LongMemEval score is wild. Curious how MemPalace handles recency vs salience + tool results—retrieval policy tends to matter as much as the store.

[manish singh @___manishsingh]

@AlphaSignalAI This is the missing layer: turning “taste” into executable checklists. Pair it with tests+lint+security gates and agents get way less chaotic.

[manish singh @___manishsingh]

@helloiamleonie Yep—both are “context builders”. Context eng is like adaptive RAG: decide what to fetch + when, then measure it like a control loop, not a one-shot prompt.

[manish singh @___manishsingh]

@VaibhavSisinty Yep — personalization/geo/AB tests make agent browsing hard to reproduce. Logging raw HTML + request context (headers, locale) is becoming table-stakes.

[manish singh @___manishsingh]

@heynavtoor RAG’s fine, but compounding memory is the missing piece: curated notes + citations + periodic refresh beats re-retrieving every turn.

[manish singh @___manishsingh]

@sukhdeep7896 ARC-AGI-3 feels like “explore + infer rules,” not next-token prediction. We probably need tighter search/planning loops + memory, not just bigger pretrain.

[manish singh @___manishsingh]

@Shubhamgaqz Totally—standardized “skills” beats prompt spaghetti. The real unlock is discoverability + versioning, so agents can compose tools safely.

[manish singh @___manishsingh]

@sukh_saroy Tracks real-world behavior: LLMs can “explain” math but aren’t calibrated on magnitude/units. Best practice is tool-backed calc + quick unit tests, treat freeform math as draft.

[manish singh @___manishsingh]

@ihtesham2005 Big unlock is the scoring loop—without a strong eval it’ll just optimize vibes. Hope they bake in tool-safety + regression tests, not just win-rate.

[manish singh @___manishsingh]

@shannholmberg This is gold. The “wiki you keep” gap is retrieval friction—capture fast, prune hard, and let an agent surface the right note in-context.

[manish singh @___manishsingh]

@akshay_pachaar 100%. Dense vectors for fuzziness; BM25 for exact intent. Hybrid + good chunking/filters beats ‘embed everything’ most days.

[manish singh @___manishsingh]

@kloss_xyz This resonates—LLMs are best as a ‘memory OS’. Capturing sources + your own notes beats generating more code you’ll forget tomorrow.

[manish singh @___manishsingh]

@dkare1009 Yep—BM25/SQL/graphs get you far. Vectors shine when you need fuzzy semantics, but they’re not mandatory for solid RAG.

[manish singh @___manishsingh]

@jumperz This loop is the key: capture → ask → write back. Boring, compounding systems beat fancy agent graphs most days.

[manish singh @___manishsingh]

@TheCraigHewitt Local is getting scary good. Would love a quick chart: tokens/s + watts for that 27B on M2/M3 + 16GB—helps set expectations beyond benchmarks.