2026-05-23

• #0024

  #0024Models & APIGeekNewslast week

  `DeepSeek-V4-Pro` Makes Discounted API Pricing Permanent

  100radar

  DeepSeek-V4-ProLLM API model — permanent one-quarter pricing

  Promo pricing is now the baseline. Token-heavy features get a cheaper model option, so cost tests deserve another pass.

  • —The 75% discount stays after the promo, keeping DeepSeek-V4-Pro at one-quarter of the prior price.
  • —Billing is based on 1M-token rates, with input and output usage deducted directly from balance.
  • —This changes the math for summarization, batch analysis, and agent loops where output volume can dominate cost.

  Source: news.hada.io/topic?id=29797Read original →
• #0023

  #0023Agents & toolsr/ClaudeAIlast week

  AI agent tool `Get Shit Done` (GSD) abandoned after creator's rug-pull scam; immediate migration required

  100radar

  Get Shit DoneAI-powered CLI automation agent — runs with local shell access

  The creator retains NPM publish rights, posing a critical security risk through malicious updates. Immediately uninstall old packages and migrate to the community fork get-shit-done-redux.

  • —The creator executed a "rug pull" with the associated $GSD crypto token and vanished. Do not trust any repos from this developer.
  • —The original NPM packages (get-shit-done-cc, @gsd-build/sdk) could receive a malicious update at any time, a high-risk vector given the tool's shell access.
  • —The community has forked the project to open-gsd/get-shit-done-redux after a security audit. Uninstall the old packages and reinstall using the new one.

  Source: www.reddit.com/r/ClaudeAI/comments/1tktl4w/if_you_use_thRead original →
• #0022

  #0022Models & APIHacker News · AI Agentlast week

  `DeepSeek V4 Pro` makes its 75% API discount permanent

  100radar

  DeepSeekLLM API provider — low-cost reasoning and coding models

  A temporary promo becomes the new baseline after May 31, 2026. Low-cost reasoning and coding API runs can now be budgeted as a standing option.

  • —The deepseek-v4-pro API price will be reset to one quarter of the original price after the promo ends on 2026-05-31 15:59 UTC.
  • —This turns a short discount window into durable infra pricing, making recurring agent workloads easier to forecast.
  • —The linked HN context mentions high caching and low cost for a native coding agent, so the price cut directly affects agent experimentation.

  Source: api-docs.deepseek.com/quick_start/pricingRead original →
• #0021

  #0021Agents & toolsClaude Code Releaseslast week

  `Claude Code` `v2.1.149` fixes sandbox and PowerShell permission gaps

  80radar

  Claude CodeTerminal coding agent — automates code work with Claude

  This is a security-heavy patch, not just polish. Workspace isolation, approval rules, and large-tree find behavior got fixes worth upgrading for.

  • —/usage now breaks limits usage down by skills, subagents, plugins, and each MCP server, making token and tool-cost debugging less blind.
  • —A PowerShell cd bypass let later commands read outside the workspace; this release closes that workspace escape path.
  • —Git worktree write allowlists now cover only shared .git, with hooks/ and config denied. Safer for repos using linked worktrees.
  • —Bash find no longer exhausts macOS file/vnode tables on large trees, reducing host-crash risk during broad repo scans.
  • —Smaller UX fixes hit /diff, task-list Markdown rendering, prompt history, transcript tailing, and slash-command hints.

  Source: github.com/anthropics/claude-code/releases/tag/v2.1.149Read original →
• #0020

  #0020Indie businessr/SaaSlast week

  `IndieAppCircle` reaches $1k earnings after 7 months of Reddit-led iteration

  70radar

  IndieAppCircleApp feedback platform — credit loop for tester exchange

  A bare MVP turned into revenue through a tight Reddit feedback loop. The useful part is the repeatable loop: ship, post, collect comments, patch fast, then monetize selected features.

  • —IndieAppCircle now has 2,841 users, 2,497 tests, and 629 uploaded apps; slow growth still compounded into paid demand.
  • —The product uses a credit loop: test other apps to earn credits, spend credits to get real users testing your own app.
  • —Distribution came from repeated Reddit posts and direct contact with commenters, not ads or a launch spike.
  • —Charging started only after feedback made the product strong enough; monetization followed usage, not the other way around.

  Source: www.reddit.com/r/SaaS/comments/1tl9i7w/after_7_months_ofRead original →
• #0019

  #0019Infra & SaaSGitHub Changeloglast week

  Staged publishing and new install-time controls for npm

  70radar

  npm enhances supply-chain security with staged publishing and new --allow-* install flags. These features provide granular control over package sources, mitigating dependency attacks.

  • —staged publishing: Publish with npm publish --stage <name> and later promote with npm stage promote. This prevents accidentally tagging a pre-release as `latest`.
  • —--allow-* flags: New flags like --allow-file and --allow-remote let you explicitly whitelist installation sources during npm install, expanding on the existing --allow-git.
  • —Both features focus on hardening defenses against supply-chain attacks. It's a good practice to configure them in .npmrc even for solo projects with many dependencies.

  Source: github.blog/changelog/2026-05-22-staged-publishing-and-nRead original →
• #0018

  #0018Generative mediafuture_tools_newslast week

  `Ideogram` Launches MCP for Design Agents in `Claude` and `ChatGPT`

  70radar

  IdeogramAI image generation SaaS — strong text rendering

  Design generation can now be called from agent workflows instead of a separate tab. Useful for quick ad creatives, app assets, and landing visuals.

  • —Ideogram exposed an MCP integration on 2026-05-22, making image/design generation callable from agent clients.
  • —Works with Claude and ChatGPT, so prompt-to-asset steps can sit inside planning, copywriting, and iteration loops.
  • —Practical fit: generate icons, promo visuals, social images, and landing-page concepts without switching tools.

  Source: x.com/ideogram_ai/status/2057863035709173993?utm_source=Read original →
• #0017

  #0017Agents & toolsGeekNewslast week

  `Zero`, an Experimental Programming Language for Agents

  60radar

  ZeroProgramming language — designed for agent-first coding

  A language designed for agent-first coding work. Deterministic inspect/repair and one obvious path per task make it worth tracking, not adopting yet.

  • —Zero comes from Vercel Labs and assumes agents, not humans, are the primary users.
  • —Deterministic Inspect and Repair target a real pain: agents need debuggable feedback loops, not clever syntax.
  • —Standard-library-first design reduces dependency choice. That can make generated code more predictable.
  • —The language is still experimental. Watch the design direction before betting product code on it.

  Source: news.hada.io/topic?id=29780Read original →
• #0016

  #0016Infra & SaaSRevenueCat Bloglast week

  RevenueCat launches `AI Toolkit` for subscription setup

  60radar

  RevenueCat AI ToolkitSubscription setup toolkit — starts setup inside coding assistants

  Subscription setup can now start inside an AI coding assistant. Useful when billing plumbing is the bottleneck, but scope is still narrow from the available details.

  • —The stated job is simple: set up subscriptions from an AI coding assistant, reducing context switching around billing setup.
  • —RevenueCat stays the billing backend; the new surface is the assistant workflow, not a separate payment provider.
  • —Available detail is thin: no pricing, supported assistants, or platform list. Treat as worth checking, not a migration trigger.

  Source: www.revenuecat.com/blog/company/ai-toolkit/Read original →
• #0015

  #0015Otherr/iOSProgramminglast week

  `swift-markdown-engine` Open-Sourced for Native macOS Markdown Editing

  50radar

  swift-markdown-enginemacOS Markdown editor engine — native TextKit 2 rendering

  A TextKit 2 Markdown editor engine is now reusable instead of buried inside one app. Useful if you need native editing; irrelevant for web-first products.

  • —Built on AppKit and TextKit 2, then bridged to SwiftUI — avoids the usual WebView Markdown editor compromise.
  • —Covers live styling, GFM tables, task checkboxes, wiki links, image embeds, LaTeX, and syntax-highlighted code blocks.
  • —Spelling and grammar skip code, LaTeX, and wiki-link tokens. That detail saves polish work in note-taking or writing apps.
  • —The author calls TextKit 2 migration rough and under-documented. Reusing the engine beats losing weekends on editor edge cases.

  Source: www.reddit.com/r/iOSProgramming/comments/1tlihof/i_opensRead original →
• #0014

  #0014Agents & toolsGeekNewslast week

  `Antigravity 2.0` Tops OpenSCAD Architectural 3D LLM Benchmark

  50radar

  AntigravityAI coding agent — Google-model-based implementation tool

  A CAD-heavy benchmark tested image-to-parametric-code work, not generic chat. Useful signal for agents that must turn visual specs into editable geometry, but still niche.

  • —OpenSCAD Pantheon uses 2 reference images plus a short prompt to test whether coding agents can produce parametric CAD code.
  • —Google Antigravity 2.0 with Gemini 3.5 Flash High scored 4.5/5, the top quality result in the benchmark.
  • —The task checks real Pantheon dimensions, inscriptions, and interior coffer details, so it rewards structured spatial reasoning over surface-level code output.

  Source: news.hada.io/topic?id=29796Read original →
• #0013

  #0013Indie businesssaastrlast week

  `Helply` Tests Pay-Only-on-AI-Resolution Support Pricing

  50radar

  HelplyAI support SaaS — charges only for resolved tickets

  The notable part is the billing model, not the mascot stunt. Free seats plus outcome-based AI fees make support software easier to trial, but revenue depends on proving real ticket deflection.

  • —Alex Turnbull previously bootstrapped Groove from $0 to $5M ARR; Helply is a second swing, not a random launch.
  • —The core offer is free forever for the support platform: no seat pricing and no usage caps in the snippet.
  • —Billing starts only when AI resolves a ticket. That aligns cost with saved support labor, but requires tight resolution tracking.

  Source: www.saastr.com/saastr-ai-app-of-the-week-helply-the-bootRead original →
• #0012

  #0012Indie businessr/SaaSlast week

  636 visitors in 28 days from SEO and Reddit with zero ad spend

  50radar

  Organic content beat paid acquisition at the earliest stage. Direct landings on the LTD page show pricing intent can be a traffic magnet, but conversion is still unresolved.

  • —Traffic came from SEO articles and a few Reddit posts, with 636 visitors in 28 days and almost all users new.
  • —About 1 in 3 visitors landed straight on the lifetime deal page from Google, skipping the broader site narrative.
  • —The weak link is not discovery but monetization: the next test is page-level conversion, offer copy, and channel attribution.

  Source: www.reddit.com/r/SaaS/comments/1tld84t/636_visitors_in_2Read original →
• #0011

  #0011Indie businessGeekNewslast week

  `Wasp` Drops Its Custom DSL After 5 Years and $5M

  50radar

  WaspFull-stack web framework — declarative apps on React and Node

  A Rails/Laravel-style web framework is moving from its own DSL to TypeScript. The lesson is blunt: language design can bury product momentum.

  • —Wasp tried a declarative app-spec language on top of React, Node.js, and Prisma; that abstraction became too costly to keep.
  • —The team spent 5 years and $5M before abandoning the custom language path. Platform bets compound slowly, then painfully.
  • —Moving to TypeScript reduces adoption friction. Familiar tooling often beats elegant syntax when distribution matters.

  Source: news.hada.io/topic?id=29778Read original →
• #0010

  #0010Idea signalsSimon Willisonlast week

  AI Data Centers Are Squeezing Cheap Device RAM

  50radar

  HBM demand is taking wafer capacity from DDR and LPDDR. Hardware costs will pressure low-end phones and device-heavy products.

  • —Only three major memory makers remain, so extra demand cannot be absorbed quickly by competition.
  • —HBM used about 2% of wafer allocation before; AI data centers could push it to 20% by end-2026.
  • —One GB of HBM consumes over 3x the wafer capacity of DDR or LPDDR, crowding out consumer RAM.
  • —Sub-$100 smartphones are already exposed. Apps targeting Africa or South Asia should keep low-RAM devices in scope.

  Source: simonwillison.net/2026/May/22/memory-shortage/#atom-everRead original →
• #0009

  #0009Idea signalsGeekNewslast week

  Memory Shortage Is Causing Repricing in Consumer Electronics

  50radar

  Skyrocketing DRAM demand from AI data centers is disrupting the consumer market. This signals rising costs for powerful local development machines, challenging the long-held expectation of ever-cheaper hardware.

  • —Demand for high-performance memory like HBM for AI servers is squeezing the supply for consumer-grade DRAM. The era of hardware's declining costs may be over.
  • —DRAM is a 'memory wall' bottleneck, improving slower than CPUs. Building new fabs costs over $15 billion and takes years, suggesting the shortage will be long-term.
  • —The 'wait-and-buy-cheaper' strategy for hardware is now questionable. It's time to re-evaluate budgets for Mac upgrades or test device purchases.

  Source: news.hada.io/topic?id=29769Read original →
• #0008

  #0008Agents & toolsGeekNewslast week

  Python 3.15's Hidden Gems: `asyncio.TaskGroup` Cancellation and More

  50radar

  Python 3.15 will make cancelling asyncio task groups cleaner, removing the need for custom exceptions. This means less boilerplate for handling async task failures, a nice quality-of-life win for future backend code.

  • —The asyncio.TaskGroup.cancel() method is being updated to work without custom exceptions, making async error handling more intuitive.
  • —This is one of the 'small but useful' changes overshadowed by headline features like the Tachyon profiler, focusing on developer ergonomics.
  • —For those using async Python frameworks like FastAPI or tools like LangChain, this reduces common boilerplate, improving the overall developer experience.

  Source: news.hada.io/topic?id=29767Read original →
• #0007

  #0007OtherGeekNewslast week

  Why Boolean Logic Matters in Legal Prompts

  40radar

  AND, OR, and NOT can turn legal prompts into explicit decision trees. Useful when building domain-specific AI flows where missing exceptions or overrequiring conditions creates bad output.

  • —AND forces every required condition to be checked, reducing missed elements in contract, policy, or compliance reviews.
  • —OR prevents optional conditions from becoming mandatory, which matters when prompts evaluate eligibility or alternatives.
  • —NOT captures exceptions, exemptions, and exclusions. That is high leverage for any checklist-style AI workflow.
  • —IF-THEN-ELSE gives a fallback path when facts or evidence are missing, instead of forcing a confident answer.

  Source: news.hada.io/topic?id=29787Read original →
• #0006

  #0006Idea signalsGeekNewslast week

  `Anna’s Archive`: bulk-accessible knowledge archive via torrents and JSON API

  40radar

  Anna’s ArchiveKnowledge archive — bulk access via torrents and JSON API

  A nonprofit archive exposes large-scale access through torrents and a JSON API, despite site CAPTCHA. Useful as a data-product signal, but copyright and load risks make commercial use fragile.

  • —Anna’s Archive aims to back up human knowledge and culture and make it globally accessible — a broad corpus signal, not just a search site.
  • —CAPTCHA protects the website from overload, while torrents and a JSON API allow bulk download. API-first access matters more than UI scraping.
  • —HTML and code are on GitLab, so the operating model is partially inspectable. Commercial reuse still needs strict copyright review.

  Source: news.hada.io/topic?id=29781Read original →
• #0005

  #0005Idea signalsGeekNewslast week

  `Gnutella`: the protocol that outlived the world that created it

  40radar

  GnutellaP2P file-sharing protocol — search and transfer without central servers

  Mass adoption came from timing, not decentralization ideology. Gnutella solved a concrete MP3 access problem when bandwidth, devices, and habits aligned. Useful as a demand-timing lesson.

  • —US internet penetration hit 50% in 2000-2001, giving P2P distribution enough reachable nodes to work.
  • —Cheaper MP3 players and weak streaming made local file ownership a practical default, not a nostalgia choice.
  • —The lesson is not P2P itself. Adoption follows when a user pain, device shift, and distribution gap line up.

  Source: news.hada.io/topic?id=29774Read original →
• #0004

  #0004Otherr/LocalLLaMAlast week

  `Qwen3.6 27B` pure `Q4_K_M` GGUF fits in **16GB VRAM**

  40radar

  Qwen3.6Open LLM — active GGUF ecosystem for local inference

  Pure quantization trims enough size to keep the whole model on a consumer GPU. Useful for local agent tests, but quality loss is real and benchmark depth is thin.

  • —Q4_K_M MTP is 15.4GB and non-MTP is 15.1GB; comparable builds listed at 16.5-18GB often spill past 16GB cards.
  • —MTP reaches 40 tok/s generation but only 195 tok/s prompt processing; non-MTP flips the trade-off at 715 tok/s pp and 24 tok/s tg.
  • —Perplexity delta is larger than Unsloth's quant: +0.1707 vs +0.0553 on MTP, so the size win buys speed/fit at some quality cost.

  Source: www.reddit.com/r/LocalLLaMA/comments/1tkzk9e/qwen36_27b_Read original →
• #0003

  #0003Agents & toolsProduct Growthlast week

  Prompt Recipe for Running Evals in `Claude Code`

  40radar

  Claude CodeTerminal coding agent — automates code work with Claude

  A compact workflow promises eval setup in minutes using exact prompts and steps. Useful as a lightweight QA loop for agentic coding, but the source gives no hard examples here.

  • —The hook is evals in minutes inside Claude Code; speed matters when shipping LLM features without a full test harness.
  • —Aparna Dhinakaran brings Arize context, so the angle is likely evaluation discipline rather than generic prompt tweaking.
  • —No metrics, sample prompts, or failure cases are included in the provided text. Treat it as a tutorial lead, not validated evidence.

  Source: www.news.aakashg.com/p/aparna-dhinakaran-podcastRead original →
• #0002

  #0002Idea signalsr/ClaudeAIlast week

  Long `Claude` Chats Still Lack In-Session Navigation

  40radar

  Search works across old chats, not inside one long work thread. The gap points to a small but sharp UX opportunity: bookmarks, anchors, and local search for AI sessions.

  • —A 40-exchange Claude session becomes hard to inspect once the answer is 30 messages back; manual scrolling is the default path.
  • —Ctrl+F does not work inside the chat pane, so normal browser search fails exactly where long work sessions need it.
  • —Asking the model to recover old context breaks down when context degrades. Navigation should be UI state, not another prompt.
  • —A browser extension or wrapper could add anchors, per-thread search, and message bookmarks without needing a new model.

  Source: www.reddit.com/r/ClaudeAI/comments/1tkto9w/claude_has_noRead original →
• #0001

  #0001Agents & toolsr/LocalLLaMAlast week

  BeeLlama v0.2.0 boosts inference speed by up to 4.9x on an RTX 3090

  40radar

  BeeLlamaLocal LLM engine — accelerates token generation via DFlash

  An inference engine that achieves up to a 4.9x token speedup over llama.cpp via DFlash. It makes high-throughput local LLMs more viable on consumer GPUs like the RTX 3090.

  • —Achieves 164 tokens/sec with Qwen 3.6 27B on a single RTX 3090, a 4.4x speedup compared to llama.cpp's 37.2 tps.
  • —DFlash, a form of speculative decoding, accelerates inference using a smaller draft model. While prompt processing speed is similar, token generation is significantly faster.
  • —The update adds full support for Gemma 4 31B and is compatible with the GGUF format, easing integration with the existing local LLM ecosystem.
  • —This makes fast prototyping or running small-scale services on owned hardware more feasible, especially for tasks involving long text generation, without cloud API costs.

  Source: www.reddit.com/r/LocalLLaMA/comments/1tkpz2y/beellama_v0Read original →