Key Takeaways
- Claude Code reverse-engineered the Battlefield 1942 demo into a browser-playable WebAssembly port, at roughly 1M output tokens and about $100 at API pricing.
- A reported GPT-6 Astra result claims direct weight edits fixed a small Qwen model's reasoning loop, drawing both enthusiasm and skepticism.
- AppSumo has cut staff from 140 to about 20 and is replacing paid software with in-house tools, including builds by non-engineers.
- MCP adoption is being questioned as agents favor APIs and browsers over manually configured connectors.
- Small-app and acquisition data points: a $400K/month iOS app, a $31.5K/month link tool, and Flippa/acquire.com listings.
- A security write-up describes roughly 24 scripted commands in 15 minutes probing credentials and internal networks, all blocked.
1. AI Agents and Model Capabilities
- Levels.io reports using Claude Code to reverse-engineer the Battlefield 1942 free trial demo and port it to browser multiplayer: disassembling x86 machine code to find 57,000 functions, translating them function-by-function into about 100,000 lines of C, compiling with Emscripten to WebAssembly, and rewriting about 300 Windows and Direct3D functions for WebGL 2/OpenGL, plus a WebSockets networking stack. The author says earlier attempts failed and this one was completed by Opus 5.5, without his knowing the exact implementation. The port consumed 1M output tokens, about $100 at API prices, though he used a $100/month Claude subscription. The server cap is 16 players and full, with attempts to raise it to 64 or 128; the Wake Island map is now playable and will join the main rotation after testing. — via 1 2 3 4
- A widely shared claim says GPT-6 Astra modified only 0.016% of Qwen3.5-0.8B's weights to fix a reasoning loop, improving math by up to 6x and instruction-following by 3x without gradients or training; the post calls this "weight-space programming." Tony Dinh remarked that a human doing this would be seen as superhuman, while Arvid Kahl compared it to Star Trek's skin regenerator, criticizing the idea of fixing a system's internals without understanding them. Patrick McKenzie proposed a term for AI progress that is fully expected yet emotionally surprising when it happens. — via 1 2 3
- Ben Tossell relayed Factory's Luke Alvoeiro saying Droid leads Terminal Bench 4 by 8 points over Anthropic's native Claude Code harness on the same Opus 5.5 model; the platform's two most-used models are Opus 5.5 and GPT-6.1 Sol, both in high-reasoning mode. The post says beating a native harness on the same model is uncommon, with a 17% cost-efficiency difference and 25% faster runs. — via 1
- On MCP, Levels.io relayed Steve Lacey's observation that AI agents rarely use MCP even when available, because connectors need manual setup, so agents call APIs or open a browser instead. The quoted author still builds MCP but frames its value as an "API without a contract," predicting connectors, directories, and install flows will die while MCP and API endpoints eventually merge with API-key and OAuth account-level access. — via 1
- Arvid Kahl flagged a security incident: about 24 scripted commands in roughly 15 minutes extracted credentials from an app container, requested all platform secrets, tried cloud metadata endpoints, probed credential permissions, guessed internal database and cache names, and scanned all 65,536 internal addresses twice. All attempts were blocked with no access; Kahl said agentic behavior would be a major red flag, and non-agentic behavior would be worse. — via 1
2. Building and Running Lean Software Businesses
- Noah Kagan says AppSumo went from 140 staff to about 20 and plans to run scheduling app TidyCal with zero people via six interlocking steps. He replaced Asana ($4,800/year) with SumoTasks, built in 3 hours by a project manager who had never coded; Teachable ($148/month) with SumoCourses, built in 2 days by a COO on a first app, now hosting 152 lessons; and Carta ($11,000/year) with SumoShares, built overnight. Two of the three replacements were by people who had never shipped an app. Kagan argues whoever feels the cost should do the replacement, and that he opposes overpaying for software that should be affordable, not paying for software. — via 1
- Alex Lieberman argues that near-zero development cost does not justify building everything, because time and cognitive-load opportunity costs are often ignored. His three build criteria: core processes or products that sustain or extend an advantage, owning your own context (organizational knowledge and evaluation data, not necessarily the model or pipeline), and major cost mismatches where replacing a vendor clears an internal ROI threshold. — via 1
- Starter Story data points: Yoni says his iOS app reaches $400K/month, after building 100K–150K Instagram followers pre-launch with teasers and a waitlist and earning about $10K on launch day. Charlie built link tool Liinks to $31.5K/month while keeping a full-time job, despite Linktree's millions of users; he says the biggest mistake is repeatedly quitting, since cheap development makes jumping to the next idea easier than doing hard things like customer contact, marketing, and outreach. A founder has bought and sold apps four times since 2022 "like flipping houses," with over $500K cumulative profit. — via 1 2 3
- Acquire.com and Flippa listings show the small-business market: Flippa added 44 First Access deals and 56 open to all users this week, including an AI care-learning SaaS at $225K, an AI-run lead-gen agency at $2.94M, a KDP romance publishing business at $976K, and a turnkey beauty brand at $928K. Acquire.com lists an accessibility e-commerce store at $2.1M TTM revenue asking $400K, a commercial real-estate SaaS at $305K TTM asking $746.9K, and a house-painter color visualization tool at $22.7K TTM asking $67.1K. It also argues "boring" software—invoicing, scheduling, inventory, compliance—is a good business, citing a plumbing scheduler with 200 customers at $250/month, or $50K MRR and $600K annual revenue, and notes acquisition offers should weigh structure, risk, and buyer certainty, not just price. — via 1 2 3 4 5 6
3. Views on AI, Work, and Money
- Codie Sanchez cites AmEx cutting 3,500 jobs, Barclays cutting 5% of staff, and Goldman Sachs saying up to 1 in 11 employees could be replaced by AI by 2040, to rebut advice to seek a stable corporate job. She also argues desire cannot be taught after hiring over 1,000 people, and that people are defined by what they finish rather than their goals. — via 1 2 3
- Nick Huber predicts longevity clinics will shift toward free, simple practices—sunlight, sunrise and sunset, circadian rhythm, dark nights, avoiding LED light on skin or eyes at night, outdoor time without sunglasses, quarterly 3-day fasts—plus bioregulatory peptides including epitalon, pinealon, vilon, vesugen, and thymulin. He also calls momentum the key to business, names an AI implementation specialist the most important hire, and says he is bullish on AI and humans but bearish on markets and valuations. — via 1 2 3 4
- Bill Perkins, relayed by Shaan Puri and Sam Parr, argues net worth peaks on a date, not at a number, and the body is a depreciating asset; he favors a "zero regrets" life over a zeroed account, lists experiences better done young, and says life has multiple "deaths" such as marriage, children, and graduation. Perkins says his goal is top 1% health for his age, noting he went from 8% body fat and top-1% VO2Max to 18% body fat and 50th-percentile VO2Max, attributing it to not doing the work rather than excuses like a newborn. — via 1 2
- Justin Welsh argues people fear spending a year working independently but will spend 50 years doing work they hate, and that large funding rounds get applause while 90%-margin solo businesses are dismissed. Tony Dinh says AI developing good taste will disrupt his career, installs Claude Code first on every server, and wants to direct tokens at useful, meaningful problems, asking how open-source projects can get frontier-level tokens. — via 1 2 3 4 5
