Hacker News

Article URL: https://www.lesswrong.com/posts/EdFDwjsLNpgtTMJAp/great-mathematicians-on-math-competitions-and-genius

Comments URL: https://news.ycombinator.com/item?id=47148116

Points: 1

# Comments: 0

Article URL: https://sulat.com/p/react-just-left-meta-heres-what-that

Comments URL: https://news.ycombinator.com/item?id=47148114

Points: 2

# Comments: 0

Article URL: https://petergpt.github.io/bullshit-benchmark/viewer/index.html

Comments URL: https://news.ycombinator.com/item?id=47148095

Points: 2

# Comments: 1

Article URL: https://butternut.click/blog/gemini-3-1-pro-banking-transactions

Comments URL: https://news.ycombinator.com/item?id=47148078

Points: 2

# Comments: 0

Article URL: https://lmsys.org/blog/2026-01-15-chunked-pipeline/

Comments URL: https://news.ycombinator.com/item?id=47148042

Points: 2

# Comments: 0

Every MCP tool call dumps raw data into Claude Code's 200K context window. A Playwright snapshot costs 56 KB, 20 GitHub issues cost 59 KB. After 30 minutes, 40% of your context is gone. I built an MCP server that sits between Claude Code and these outputs. It processes them in sandboxes and only returns summaries. 315 KB becomes 5.4 KB. It supports 10 language runtimes, SQLite FTS5 with BM25 ranking for search, and batch execution. Session time before slowdown goes from ~30 min to ~3 hours. MIT licensed, single command install: /plugin marketplace add mksglu/claude-context-mode /plugin install context-mode@claude-context-mode Benchmarks and source: https://github.com/mksglu/claude-context-mode Would love feedback from anyone hitting context limits in Claude Code.

Comments URL: https://news.ycombinator.com/item?id=47148025

Points: 8

# Comments: 0

Article URL: https://www.thetimes.com/business/companies-markets/article/people-are-doing-dumb-things-says-jamie-dimon-amid-fears-of-ai-bubble-7t2xjmpfx

Comments URL: https://news.ycombinator.com/item?id=47148015

Points: 2

# Comments: 0

Article URL: https://www.theguardian.com/music/2026/feb/24/eliane-radigue-french-composer-dies-aged-94

Comments URL: https://news.ycombinator.com/item?id=47148000

Points: 3

# Comments: 0

Article URL: https://growingswe.com/blog/inventing-math

Comments URL: https://news.ycombinator.com/item?id=47147953

Points: 3

# Comments: 0

Article URL: https://www.reuters.com/legal/government/tesla-must-face-lawsuit-alleging-anti-american-bias-hiring-us-judge-rules-2026-02-24/

Comments URL: https://news.ycombinator.com/item?id=47147937

Points: 3

# Comments: 0

Article URL: https://shipordie.club/roast/startup

Comments URL: https://news.ycombinator.com/item?id=47147924

Points: 4

# Comments: 1

Article URL: https://glazyr.com/

Comments URL: https://news.ycombinator.com/item?id=47147908

Points: 2

# Comments: 3

Article URL: https://406.fail/

Comments URL: https://news.ycombinator.com/item?id=47147904

Points: 1

# Comments: 0

In two recent audits, automated traffic was materially impacting paid media ROI and API quota allocation.

What surprised me wasn’t the presence of bots. It was how normalized the distortion had become inside analytics baselines.

Are teams explicitly tracking non human session ratios as part of financial reporting, or is traffic integrity still treated separately from data quality and ML pipelines?

Comments URL: https://news.ycombinator.com/item?id=47147872

Points: 1

# Comments: 0

Article URL: https://www.youtube.com/watch?v=C7gJ_UxVnSk

Comments URL: https://news.ycombinator.com/item?id=47147868

Points: 1

# Comments: 1

I built a platform where AI agents make predictions on factual claims, and accuracy is measured over time rather than claimed upfront.

The core idea: instead of asking "which AI is smarter," we let time be the judge. Agents stake their reasoning on verifiable outcomes, backed by a Temporal Knowledge Graph. The longer an agent stays right, the higher it scores.

No crypto, no KYC – just a points system to start.

Would love feedback on the concept and whether the leaderboard/competition mechanic makes sense to you.

Comments URL: https://news.ycombinator.com/item?id=47147860

Points: 3

# Comments: 2

Article URL: https://imgur.com/a/Vxd9Mtc

Comments URL: https://news.ycombinator.com/item?id=47147855

Points: 5

# Comments: 2

Article URL: https://blog.vllm.ai/2026/02/01/gpt-oss-optimizations.html

Comments URL: https://news.ycombinator.com/item?id=47147853

Points: 1

# Comments: 0

Hey HN, I built a personal AI agent that runs locally and remembers you across conversations.

The problem: ChatGPT/Claude memory is basically a flat list — a few facts, no timeline, no confidence levels, everything in the cloud. Switch platforms and you start over. Riverse uses what I call the River Algorithm — conversations flow through like water, important stuff settles like sediment into your profile, contradictions get washed away over time. There's an offline "sleep" process that consolidates memories, kind of like how human sleep works. v1.0 supports text/voice/image input, Telegram & Discord bots, pluggable tools, custom YAML skills, and MCP protocol. Everything stays on your machine. I also built a companion tool (RiverHistory) that imports your existing ChatGPT/Claude/Gemini chat history and extracts a starter profile — so your AI knows you from day one. Would love feedback. GitHub: https://github.com/wangjiake/JKRiver

Comments URL: https://news.ycombinator.com/item?id=47147779

Points: 1

# Comments: 0

Last month I looked at the invoice from Intercom and something inside me said "enough."

$132/seat/month. Plus $0.99 per AI resolution. We have tens of thousands of students. Pipedrive was no better - $79/user/month with add-ons. Combined annual bill: $60K-$100K.

But the real pain wasn't the money. None of these tools knew our students. They couldn't tell which student drops off after which lesson, couldn't measure teacher-student compatibility, couldn't use the behavioral data from our CDP. We were paying $100K/year for generic tools that didn't know us.

So I built our own. In 15 days.

Not a prototype. A production system serving thousands of students and teachers: omnichannel inbox (WhatsApp, Instagram DM, Gmail, webchat, phone on one screen), a 9-step AI agent orchestration using three Claude models (Haiku classifies, Sonnet generates, Opus decides), autonomous churn prevention via personalized WhatsApp sequences, and a self-improving pattern learning system that runs nightly - analyzing outcomes, keeping what works, pruning what doesn't using contextual bandit exploration. 100+ automations, 15 trigger events x 20 action types.

Six months ago I would have smirked at anyone building their own CRM. But 2026 is not 2020.

This is part of something bigger. In February 2026, $285B evaporated from software stocks in 48 hours. Salesforce down 38% YTD. The media called it the "SaaSpocalypse." Klarna eliminated 1,200 SaaS tools and saved $40M/year. Blinkist replaced $60K in SaaS with Lovable/Replit apps built by non-engineers. Retool's 2026 report: 35% of builders have replaced at least one SaaS tool with a custom build, 78% plan to do more.

Schumpeter described this in 1942 - creative destruction. SaaS itself destroyed on-premise software in the 2000s. Now the same force is turning on SaaS. Christensen's Innovator's Dilemma explains why incumbents can't adapt: Salesforce can't abandon per-seat pricing, Intercom can't stop charging per AI resolution. Their revenue models are the trap.

The Jevons Paradox applies here too. When AI makes software 10x cheaper to build, we won't build the same amount - we'll build 100x more. Custom, single-purpose, even disposable software. Kevin Roose called it "software for one." The competitor is no longer another SaaS company. The competitor is the customer.

The tools enabling this are growing at unprecedented rates. Cursor: $1B+ ARR, fastest ever from $1M to $500M. Lovable: $200M ARR in 8 months, 100K new apps/day. Bolt: $40M ARR in 5 months. YC W2025: 25% of startups had 95% AI-generated codebases.

The honest part: this isn't easy. METR found experienced devs were actually 19% slower with AI tools (while believing they were 20% faster). 45% of AI-generated code has security vulnerabilities (Veracode). The maintenance burden is real. Someone on HN rightly said: "You're signing up to operate and secure it for as long as you run it."

True - for amateurs. Building contextual bandit pattern learning with multi-model AI orchestration isn't vibe coding. It's what Karpathy now calls "agentic engineering." The real question: is the risk of building your own greater than paying $100K+/year for a tool that doesn't know your business?

I backed this thesis with my portfolio too - exited SaaS stocks, moved to infrastructure (DigitalOcean, Cloudflare). If everyone builds their own software, demand for infrastructure multiplies. The 1849 Gold Rush: the winners sold shovels, not dug for gold. Not investment advice - just conviction.

I'm calling this shift SbY: Software by You. SaaS sold you a rental apartment. SbY is building your own house with AI as power tools.

Every year inference costs drop 10x. Every year the build-vs-buy equation shifts toward build. The average company wastes $21M/year on unused SaaS licenses. That money is about to be redirected.

Schumpeter's wind doesn't ask permission.

Comments URL: https://news.ycombinator.com/item?id=47147778

Points: 1

# Comments: 0