Top 9 Hacker News posts, summarized
HN discussion
(495 points, 296 comments)
Fields Medalist Terence Tao engaged in an extended ChatGPT conversation exploring a purported counterexample to the Jacobian Conjecture, a long-standing open problem in polynomial mapping theory. The dialogue shows Tao probing the mathematical structure of the counterexample—specifically a polynomial map with a Jacobian determinant identically equal to 1 that nonetheless fails to be invertible—requesting geometric explanations, coordinate-free interpretations, and verification of algebraic properties. Tao steers the conversation with precise technical prompts, iteratively refining representations and decomposing the problem to understand why the counterexample works, demonstrating a collaborative workflow where the LLM performs symbolic computation, generates proofs, and suggests alternative framings while Tao provides mathematical direction and validation.
Commenters emphasize the sophistication of Tao's prompting strategy: short, jargon-dense queries that steer the model rather than open-ended questions, causing the LLM to match his technical level. Observers note the model's ability to push back with caveats and corrections, maintain context over a long multi-modal session (code, PDF ingestion, reasoning), and verify computations that contradict its training cutoff knowledge—creating a "cognitohazard" where the LLM can formally verify a counterexample it "knows" shouldn't exist. Many highlight the broader implication: this interaction exemplifies a future of AI as genuine research copilot in elite technical domains, where the expert's domain knowledge unlocks dramatically more value from the same model. Programmers draw parallels to software modeling workflows—observe, re-represent, decompose, verify, re-integrate—while non-specialists express both awe at the inaccessibility of modern mathematics and anticipation for "explainer agents" that could bridge the gap.
HN discussion
(404 points, 113 comments)
John C. Dvorak, longtime technology columnist and co-host of the No Agenda Show podcast, has died at age 74. The brief announcement describes him as a "beloved husband, father, and cherished host" and indicates more details will follow. Dvorak was a prominent figure in tech journalism for decades, known for his columns in PC Magazine, InfoWorld, and Byte, as well as appearances on ZDTV/TechTV and the TWiT network.
Commenters uniformly expressed sadness and nostalgia, with many recalling Dvorak's columns as essential reading during the 1980s and 1990s PC era. Several noted his distinctive cranky, contrarian style—one user referenced his "bad Apple predictions" while another linked to his recent PC Magazine opinion pieces on topics like AI hype cycles and 3D printing. Multiple commenters clarified he was not the inventor of the Dvorak keyboard (that was his uncle, August Dvorak). His small thumbnail photo in PC Magazine was cited as an iconic memory. There was minor confusion about his age (Wikipedia lists 80 vs. the announcement's 74). Several referenced his later work on Cranky Geeks and No Agenda, with one commenter suggesting the news warranted a "black bar" tribute on Hacker News.
HN discussion
(246 points, 215 comments)
The article examines whether creatine supplementation improves cognitive function. It establishes creatine as a safe, well-studied nutrient that reliably increases muscular strength and power (5–15%) by expanding phosphocreatine energy reserves for short bursts. The brain also contains creatine, and genetic deficiencies cause intellectual disability, confirming a critical neurological role. Supplementation modestly raises brain creatine levels (3–10%), but the mechanistic link to enhanced cognition remains unclear—creatine appears to shuttle energy locally in neurons, yet the brain lacks the burst-energy demands that make creatine effective in muscle.
The evidence for cognitive benefits is weak and inconsistent. A seminal 2003 study (Rae et al.) reported massive IQ gains (~17–26 points) in vegetarians, but a high-quality 2023 replication (Sandkühler et al., n=123) found negligible effects (~1–2.5 IQ points, non-significant) and no vegetarian-specific benefit. Meta-analyses suffer from methodological flaws (double-counting data), and the European Food Safety Authority’s 2024 review concluded no consistent cognitive effect across domains. The author estimates a small possible benefit (0.1–0.3 SD) primarily under metabolic stress (sleep deprivation, aging, vegetarian diet), but considers the expected value (~0.5 IQ points) too small to justify daily supplementation for cognition alone.
Commenters largely align with the article’s skepticism. Several report subjective benefits only during sleep deprivation or mental fatigue (headache reduction, feeling “more energized to think”), not baseline enhancement. One argues the null result should be interpreted as evidence of no effect given the low prior probability of supplement efficacy and pervasive publication bias in nootropic research. Others attribute perceived benefits to increased hydration from creatine-induced water retention, while a few cite Gwern’s prior analysis supporting strength benefits but not cognition. A commenter disputes the article’s claim that intense thinking doesn’t increase caloric expenditure, citing contrary evidence.
Practical questions dominate: cycling necessity, pill vs. powder formats, pediatric safety, and drivers of recent popularity. One commenter flags a potential dosage typo in the cited ISSN position stand (0.3–0.8 g/kg/day). Historical context notes past supplement contamination with prohormones skewed risk/benefit perceptions. Caffeine is identified as the only other Tier-1 evidence-backed strength supplement.
HN discussion
(326 points, 131 comments)
The article investigates whether AI labs are "pelicanmaxxing" — specifically optimizing models for the popular "pelican riding a bicycle" SVG benchmark created by Simon Willison. The author tested seven frontier models (GPT-5.6 Terra, Claude Sonnet 5, Gemini 3.5 Flash, Grok 4.5, Qwen3.7-Max, GLM-5.2, DeepSeek V4 Pro) across 48 animal-vehicle combinations (8 animals × 6 vehicles), generating 1,008 SVGs total. Using an LLM judge (GPT-5.6 Luna) and feature extraction (Gemini 3.1 Flash-Lite), the analysis examined five lines of evidence: visual inspection, animal rankings, vehicle rankings, regression analysis controlling for inherent difficulty, and scene composition analysis. Results show pelicans rank 6th of 8 animals, bicycles rank 2nd-to-last of 6 vehicles, and the pelican-bicycle combination ranks 42nd of 48. Fixed-effects regression found no statistically significant per-lab boost for pelicans, bicycles, or the specific pelican-bicycle cell after multiple-comparison correction (only Gemini showed a marginal bicycle effect consistent with a false positive). All 21 pelican-bicycle images faced right, but this matches general tendencies (60% of all images face right; bicycles and pelicans independently favor right-facing). Scene elements showed no memorization pattern unique to pelican-bicycle. The author concludes there is little evidence of targeted benchmark gaming, suggesting instead that labs may be "SVGmaxxing" — broadly improving SVG generation capability. Limitations include a single judge, inability to detect general SVG optimization, and limited sample size (3 per cell, ~$80 budget).
Commenters largely praised the rigorous methodology, with Simon Willison himself calling it "fantastic" and more robust than his own spot-checks. A key technical insight came from mauvehaus, who explained the universal right-facing bias: bicycle drivetrains are on the right side, so training data predominantly shows right-side views, making this a data artifact rather than memorization. Multiple commenters distinguished between specific benchmark gaming ("pelicanmaxxing") and general SVG capability improvement ("SVGmaxxing"), with Wowfunhappy noting the latter is a genuinely useful skill. andy99 argued that if labs were gaming, they would use data augmentation across animal-vehicle pairs during training rather than fine-tuning on a single prompt. apwheele contrasted the benchmark with real-world SVG tasks (e.g., simple icons), where models struggle despite pelican success, highlighting the benchmark's limited ecological validity. tomas789 suggested pairwise Elo ratings would be more reliable than absolute scoring. Several commenters noted specialized SVG models (e.g., Refract V4) outperform general LLMs significantly. andrewstuart criticized the benchmark as absurd and unrepresentative, comparing it to notorious Microsoft interview questions. BeetleB humorously noted that publishing the dataset makes it easier for future training contamination. Overall, the discussion validated the article's negative finding while debating the benchmark's value and proposing better evaluation approaches.
HN discussion
(286 points, 161 comments)
The article provides a practical guide to running PostgreSQL in production at scale, drawn from the author's experience at Hatchet. It progresses from fundamentals—schema design (identity columns, `timestamptz`, primary/foreign keys), efficient read queries (indexes vs. sequential scans, compound indexes aligned with `ORDER BY`), write patterns (short transactions, `CREATE INDEX CONCURRENTLY`), and connection pooling (pgbouncer or in-memory pools)—to intermediate topics like the query planner (statistics, `EXPLAIN ANALYZE`, when to accept seq scans), bulk inserts (batching for 10× throughput), and autovacuum tuning (dead tuples, transaction ID wraparound, table/index bloat with `pg_repack`/`REINDEX CONCURRENTLY`). Advanced sections cover `FOR UPDATE SKIP LOCKED` for queue/lease patterns, partitioning for time-series data (independent autovacuum, instant partition drops), and large-table migration techniques using triggers and batched backfills outside transactions.
Commenters identified several critical omissions: monitoring/alerting for failure modes like XID wraparound (thundergolfer), backup/restore strategy (theallan, citing Barman), and stored functions for SQL injection prevention (traceroute66). Multiple voices emphasized schema discipline—normalization fundamentals (zer00eyz recommending *Database Design for Mere Mortals*), avoiding ORM-generated schemas, and using `uuidv7`/`timestamp` over `timestamptz` (lennoff, ComputerGuru). Technical corrections included hash/GIN/BRIN indexes, deterministic lock ordering to prevent deadlocks, `EXPLAIN (GENERIC_PLAN)`, and histograms for skewed data (ComputerGuru, sgarland). Practical debates surfaced around cascading deletes (mjr00 prefers explicit application-level deletes), in-memory joins vs. complex SQL (mrkaye97), and JSONB denormalization trade-offs (tracker1, sgarland). Tooling recommendations included pg-osc for online schema changes, Flyway/Liquibase for migration management, and Grate for .NET. A recurring theme was organizational simplicity: avoid ORMs, use append-only sources of truth, minimize explicit transactions/locking, and don't reinvent graph/type systems in SQL (frollogaston). Cost-conscious founders noted managed Postgres can exceed $100/mo, pushing some toward SQLite/DuckDB/DynamoDB (hmokiguess).
HN discussion
(194 points, 193 comments)
The author criticizes Reddit's decision to require login for accessing old.reddit.com, which Reddit claims is for "safety" against abusive scraping. The author argues this is a pretext to protect Reddit's valuable user-generated content for AI licensing deals, noting that old.reddit.com serves plain HTML (lightweight, fast) while new Reddit requires JavaScript and loads 5x more data. The author demonstrates that new Reddit's "security" appears to rely on browser fingerprinting and JavaScript execution rather than actual content protection. They express frustration at being forced to authenticate to access human-written knowledge that was previously publicly accessible, and predict logged-out access to new Reddit will eventually be restricted as well.
Commenters largely agree the move is driven by Reddit's AI licensing deals with OpenAI and Google to exclude competing scrapers, with several noting this represents a "second death" of Reddit after its 2015 VC pivot. Many long-time users state they will abandon the platform entirely since old.reddit.com was their only usable interface. Some defend Reddit, citing genuine scraper overload and the maintenance burden of supporting two frontend stacks. Multiple comments document declining content quality (bots, astroturfing, moderator abuse) and discuss alternatives like Redlib (a Nitter-like frontend), self-hosted Reddit archives, or moving to platforms like Mastodon. A few note workarounds like narrowing the browser window to bypass the login prompt.
HN discussion
(305 points, 56 comments)
GigaToken is a new Rust-based tokenizer claiming ~1000x speedup over HuggingFace's tokenizers and tiktoken, with drop-in compatibility for both APIs. It supports nearly all common tokenizers (BPE-based models like Llama, Qwen, DeepSeek, Gemma, etc.) and runs on modern x86 and ARM CPUs. Benchmarks on an 11.9 GB OpenWebText dataset show throughput of 8–24 GB/s depending on hardware (Apple M4 Max: 8.3 GB/s; dual-socket AMD EPYC 9565: 24.5 GB/s), compared to 6–25 MB/s for HuggingFace. The native GigaToken API achieves the highest speeds by reading files directly in Rust with maximum parallelism, while compatibility mode (wrapping HF/tiktoken tokenizers) remains significantly faster but below the 1000x mark. Key optimizations include SIMD-accelerated pretokenization (replacing regex engines), aggressive caching of pretoken mappings, minimized Python-Rust boundary crossings, and reduced inter-thread communication. Limitations include slower SentencePiece support, no WordPiece support yet, limited Windows testing, and no file sinks in the native API. The author estimates tokenizing all of Common Crawl (130T tokens) would take ~6.5 hours on the EPYC system.
Commenters expressed amazement at the benchmark numbers but questioned practical impact, noting tokenization typically consumes <0.1% of inference time. Several highlighted that the real value lies in offline pre-training data preparation—tokenizing terabytes of corpus data—where speed directly reduces iteration cycle time and compute costs. One user asked about compatibility-mode benchmarks (not provided in the article). Others debated whether similar 1000x opportunities exist elsewhere in the inference pipeline, while a few framed the effort as "over-engineering" for a minor runtime component. The author clarified that optimizations generalize across CPUs and tokenizers, not overfit to specific hardware. Adoption questions arose: users would need to integrate GigaToken into their own pipelines, as HF models don't automatically use it.
HN discussion
(245 points, 102 comments)
Brian "Beej" Hall, a veteran developer and CS instructor, explores the psychological distinction between *making* things oneself and *commissioning* them from AI or other people. After presenting AI-generated creative work (fiction, artwork descriptions, game code) as his own, he reveals the deception to illustrate his discomfort: he cannot claim authorship of output he merely prompted, just as he would say "I had a deck built" rather than "I built a deck." He contrasts this with tools like compilers or hammers, which feel like extensions of his own agency because their behavior is mathematically precise and predictable. Writing a modest flashcard app by hand (177 lines) gave him infinitely more pride than any AI-generated artifact, despite taking 50x longer. Hall acknowledges the skill in prompting—vision, judgment, communication—but argues it constitutes *management*, not *making*. The essay ultimately grapples with a fuzzy but deeply felt boundary: the satisfaction of craft comes from the understanding gained through direct implementation, not from the mere existence of a finished product.
Commenters largely resonated with Hall's distinction, framing it as a question of *understanding* versus *delegation*. Several noted that AI-assisted work leaves no durable mental model—unlike hand-written codebases they can still navigate years later—while others compared AI to CNC machines or 3D printers: powerful enablers that shift the craft to a different abstraction layer. A strong minority (e.g., planb, bredren) rejected the premise, arguing that pride in outcomes (a custom guitar-tab editor, a landscaped garden) need not depend on manual implementation, and that "vibe coding" demands its own skills—taste, architecture, tooling—that constitute a legitimate modern craft. The compiler analogy drew scrutiny: layer8 emphasized that compilers are predictable and debuggable at the source level, whereas LLM output is opaque and non-deterministic. Others highlighted practical value—AI as a "software 3D printer" for niche tooling (Multiplayer's ATEM controller) or unlocking previously impossible solo projects (guybedo). A recurring theme was the erosion of side-project motivation (hbcdbff, ge96) and a desire for social norms to distinguish AI-generated from human-made work (sashank_1509).
HN discussion
(171 points, 146 comments)
The author describes returning to Kagi after experimenting with alternatives including Google, DuckDuckGo, Brave Search, Qwant, and a self-hosted SearxNG instance. They found Google's results degraded with excessive AI, video, and image content, while SearxNG suffered from rate-limiting and inconsistent quality. Other engines failed to match Kagi's result relevance, summarization, translation features, and CSS customization capabilities. The author concludes that Kagi's privacy-focused approach and superior text-centric results make it irreplaceable, and they cannot return to engines that prioritize AI-driven multimedia over textual search.
Many commenters express strong satisfaction with Kagi, citing features like vim keybindings, explicit AI opt-in, domain blocking/boosting, and superior translation as reasons for loyalty. Long-term subscribers feel the service aligns with user interests because its subscription model avoids ads and sponsored results. However, a significant theme is the impact of LLMs: several users report drastically reduced search volume as they shift to AI tools, leading some to downgrade or question the subscription's value. Others criticize the $10/month price as too high, note the limited 300-search $5 tier, and request features like API access, better international product search, Safari integration, and widgets. A few compare Kagi favorably to Brave Search's independent index and anonymity, while one commenter observes that the web's declining content quality, not Kagi, makes search feel worse than a decade ago.
Generated with hn-summaries