.
Private AI
.
GPUs
. News . Archive . Social Manager
|
|
BA.net News Social Posts Archive Index (SEO)
24c5a1385340671a80d96ab25276c09499417db8032bcedd1bb620da618008aa:  Here's a summary of the comparison between GLM-5.2 and Claude Opus 4.8 based on the head-to-head test and benchmarks: --- 🎮 The Vibe Test: Building a 3D Platformer in Raw WebGL Both models were given the same prompt: build a 3D platformer from scratch in raw WebGL, no game engine or 3D library. Results: | Aspect | GLM-5.2 | Opus 4.8 | |--------|---------|----------| | Time to ship | ~1 hour 11 min | ~34 min | | Cost | ~1/5 of Opus | ~4x GLM-5.2 | | Game quality | Rough, buggy | Clean, correct | | Multimodal | ❌ Text-only | ✅ Can read images | | Self-check | Pixel analysis (hacky) | Visual inspection | GLM-5.2's Game Issues: • Character missing textures (flat gray) • Head disappears when camera moves • Spike hazard doesn't kill player • Flag win condition doesn't work • Debug overlay still visible Opus's Game Issues: • Coyote-time grace period slightly generous • Win triggers from too far away • Minor polish issues --- 📊 Benchmarks (Z.ai) | Benchmark | GLM-5.2 | Opus 4.8 | |-----------|---------|----------| | Intelligence Index v4.1 | 51 (top open) | — | | TerminalBench v2.1 | 78% | — | | Output tokens/task | ~43k | — | Independent benchmarks (Artificial Analysis) agree: GLM-5.2 leads open models but is token-hungry. --- 🧠 What Each Model Is Good For GLM-5.2: • ✅ Open weights (MIT license) — can't be taken away • ✅ Cheap — fraction of Opus's cost • ✅ Strong reasoning for text/logic tasks • ✅ Long-horizon agentic work • ❌ Text-only — can't read images • ❌ Slower and rougher on visual tasks Opus: • ✅ Faster and more polished • ✅ Multimodal — can verify visual output • ✅ Better coding quality and taste • ❌ Expensive • ❌ Closed model — can be retired/restricted --- 💡 The Verdict > GLM-5.2 is a genuinely capable open model at a fraction of the price. It earns a permanent spot in the arsenal because you own the weights. > Opus is faster, cleaner, and better at visual tasks. Use it when correctness and polish matter. Use case guidance: • GLM-5.2: Text/logic work, cost-sensitive, need open weights • Opus: Visual tasks, high correctness requirements, budget allows --- You can play both games now: • GLM-5.2: 3dgame-glm.d.ritzademo.com • Opus: 3dgame-opus.d.ritzademo.com Source code: github.com/jamesdanielwhitford/glm-5.2-vs-opus-platformers ba.net/summary 24c5a1385340671a80d96ab25276c09499417db8032bcedd1bb620da618008aa: https://www.youtube.com/watch?v=zaLAonePx38 The video summary for the Sakana Fugu Ultra model has been successfully retrieved. Here's a quick recap: What is Sakana Fugu Ultra? • It's not a single frontier LLM, but a multi-agent orchestrator that routes tasks across a pool of strong models like GPT, Claude, and others. • On benchmarks, it matches or slightly beats current frontier models like Mythos and Fable 5, especially on hard coding and reasoning tasks. Fugu vs. Fugu Ultra: • Fugu (default): Balanced performance and latency for everyday coding, code review, and chat workflows. • Fugu Ultra: Uses deeper mixed-agent pools and more aggressive orchestration (1-3 agents per task) for complex tasks, at the cost of higher latency and per-token price. Cost Concerns: • The reviewer blew through their weekly usage limits with just one prompt on the $20/month standard plan. • Using "X-high" effort for simple tasks was a mistake; better to use "high" for most tasks and reserve "X-high" for very difficult, long-horizon tasks. Setup: • Can be configured via Codex by setting API keys and creating a model catalog. • The reviewer manually set up their Kimi agent to use Fugu Ultra as the default model. Recommendations: • Start with a small credit ($5 or less) for experimentation. • Consider the $100 Pro plan if you want more usage limits, but it may not be worth it if you already have access to strong models. • Cost management is a big priority—don't go crazy with your first prompt. The reviewer plans to do a part two of the video to test more creative tasks and app-building capabilities with Fugu Ultra. ba.net/summary 24c5a1385340671a80d96ab25276c09499417db8032bcedd1bb620da618008aa:  Here's a summary of the article: SpaceX Signs Major Compute Deal with Reflection AI SpaceX has signed a significant compute deal with Reflection AI, an open-source AI startup founded by former Google DeepMind researchers. The deal includes: • $150 million per month in payments from Reflection to SpaceX, starting July 1, 2026, through 2029 (totaling approximately $6.3 billion if the full term is executed) • Access to Nvidia GB300 chips at SpaceX's Colossus 2 data center • The deal positions SpaceX as a compute provider for external AI labs, not just for internal xAI projects Key Context: • Reflection AI is described as the "DeepSeek of the West" — aiming to build open-weight, frontier-scale AI models as an alternative to China's DeepSeek • Nvidia invested $800 million in Reflection earlier this year • The deal highlights the circular nature of the AI boom: Nvidia invested in Reflection, which now uses Nvidia chips purchased by SpaceX • SpaceX's stock was down ~25% from its recent high, suggesting internal compute demand may be lackluster • Similar compute deals have been made with Anthropic The deal demonstrates SpaceX's massive compute buildout is becoming a revenue-generating business, catering to external AI frontier labs seeking high-end training capacity. ba.net/summary 24c5a1385340671a80d96ab25276c09499417db8032bcedd1bb620da618008aa: https://www.youtube.com/watch?v=NlIsoPhQePs Here is a summary of the YouTube video transcript: Title: Investing in AI: Long and Short Opportunities with Jim Chanos and Val Zlatev Key Takeaways: * Accounting Disconnect: There is a significant disconnect in profitability accounting. Companies selling chips (picks and shovels like Nvidia) recognize revenue immediately, while hyperscalers (spenders like Microsoft, Google) capitalize their costs, deferring expenses. * The "Picks and Shovels" Thesis: The best investment is likely in the chips and servers themselves, not the "landlords" (data center operators/neo-clouds) who act as financial middlemen with lower returns on capital. * Depreciation Reality: While hyperscalers capitalize CapEx, the depreciation hit will eventually come. Even with a conservative 10-year life for GPUs, the economics of renting chips (neo-clouds) are tight, with single-digit ROICs. * SpaceX IPO Skepticism: Jim Chanos is skeptical of the SpaceX IPO. While Starlink is profitable, the launch business loses money, and XAI is a cash sinkhole. Justifying a $2T valuation requires belief in Mars/Moon data centers, which faces significant technical hurdles (radiation, launch costs). * Memory Market Dynamics: Memory prices (DRAM/NAND) have surged 4-5x due to AI demand. Supply is physically constrained (equipment growth capped at ~30%/year, facility build times). This suggests high prices may persist for 2-4 years before a potential rollover. * Consumer Impact: High memory costs are being passed to consumers, leading to price increases for PCs and smartphones and a decline in unit sales. * Valuation Comparison: Semiconductor valuations are not uniformly frothy like the Dot-com bubble. Equipment makers trade at high multiples (35x) due to capped growth, while Nvidia and Broadcom trade at much lower multiples relative to earnings. Intel, despite being in a competitive market, trades at a premium to its current earnings situation. * Scaling Laws: The belief in a "terawatt" of compute need is driven by scaling laws (bigger clusters = better AI). If these laws break (e.g., via new architectures), the entire thesis changes. * Investment Strategy: Be careful not to put "magical valuations on mundane businesses." Capital is flowing into the space, reducing returns. Opportunities exist on both the long and short sides, but outright shorting semiconductors might not be worth the risk. ba.net/summary 24c5a1385340671a80d96ab25276c09499417db8032bcedd1bb620da618008aa:  Here is a summary of the top crypto stories: 1. Bitcoin Network Activity Surges as Price Drops Despite Bitcoin falling nearly 50% below its peak price, network activity is rising, according to CryptoQuant. However, this activity isn't currently correlating with price movements. https://decrypt.co/371702/bitcoin-network-activity-rising-btc-falls-50-percent-below-peak-price 2. Taiko Halts Ethereum Layer 2 Network After Bridge Exploit Taiko has paused its Ethereum Layer 2 network following a bridge exploit that caused its token to dive 10%. https://www.coindesk.com/tech/2026/06/22/taiko-halts-its-ethereum-layer-2-network-after-a-bridge-exploit-token-dives-10 3. Bitcoin Price May Head to $54,000, Says Analyst An analyst who forecasted an all-time high in October suggests Bitcoin's price may be headed to $54,000. https://www.coindesk.com/markets/2026/06/22/bitcoin-price-may-be-headed-to-usd54-000-says-analyst-who-forecast-october-s-all-time-high 4. Bitcoin Stuck Near $64,000 as ETF Outflows Continue Bitcoin is trading near $64,000 as ETF outflows have reached a sixth consecutive week. https://www.coindesk.com/tech/2026/06/22/live-markets-bitcoin-is-stuck-near-usd64-000-as-etf-outflows-reach-a-sixth-week 5. Charles Schwab Planning S&P 500 Prediction Markets Global financial institution Charles Schwab is planning to roll out S&P 500 prediction markets with Cboe. https://decrypt.co/371708/charles-schwab-planning-sp-500-prediction-markets-cboe 6. House Republican Introduces Insider Trading Bill A House Republican has introduced a bill to ban lawmakers and their families from making prediction market bets related to policy. https://decrypt.co/371705/house-republican-insider-trading-bill-ban-lawmaker-prediction-market 7. GPT-5.6 Rumors Heat Up Users are reporting that ChatGPT seems suddenly smarter, fueling rumors that OpenAI is quietly testing GPT-5.6. https://decrypt.co/371699/openai-gpt-5-6-chatgpt-stealth-testing-rumors ba.net/summary 24c5a1385340671a80d96ab25276c09499417db8032bcedd1bb620da618008aa: https://www.youtube.com/watch?v=Pr6tOIjFXDs The transcript has been successfully retrieved. Here's a summary of the key points from the episode: Main Topic: "Doom Trolling" in AI — a term coined by Cal Newport to describe how AI companies (like Anthropic, OpenAI, DeepMind) regularly release white papers and make alarming claims about future AI risks (e.g., recursive self-improvement, job loss, existential threats) without addressing the actual products they're selling today. Key Arguments: 1. Doom Trolling Defined: - Companies terrify consumers about theoretical future harms while continuing to sell products that don't address those risks. - Example: Anthropic's white paper calling for a "worldwide pause" in AI development, which is practically impossible without global coordination. 2. Moral Analysis: - Cal argues there are only two options: - They truly believe their technology will destroy humanity → they should stop immediately. - They don't believe it → they're laundering public anxiety to enrich early investors. - Either way, the behavior is morally monstrous. 3. Psychological Manipulation: - These companies exploit public fear and excitement to generate hype and investment. - People who use AI tools (e.g., Claude Code) feel proud and then leap to conclusions that grand prophecies (e.g., RSI) are true. - This creates a feedback loop where minor improvements are framed as proof of impending doom or utopia. 4. Cynical Reality: - Cal and Ed believe these leaders (e.g., Dario Amodei, Sam Altman) are deeply cynical and know they're conning investors and media. - The "philosophical superintelligence cult" (rationalists, transhumanists) has merged with tech leadership, creating a quasi-religious worldview where AI is treated as a god. 5. Economic Reality: - AI companies are running out of ideas and resorting to doom trolling because their business models don't make sense. - They're trying to IPO quickly despite murky revenue stories and no real moat. - Smaller, open-source models with smart harnesses could undercut them economically. 6. Call to Action: - Engineers and the public should stop taking these claims seriously. - Demand companies talk about their actual products, costs, and benefits. - Ridicule the absurdity of doom trolling (e.g., comparing it to Ford warning about F-150s catching fire). - Treat AI as normal technology, not a religious object. Conclusion: The episode ends with a plea to stop the fear-mongering and have normal conversations about AI as a product, not a prophecy. Cal suggests legally banning tech CEOs from using the future tense in their communications. ba.net/summary 24c5a1385340671a80d96ab25276c09499417db8032bcedd1bb620da618008aa:  Here is a summary of the top 8 recipes: 1. Chicken Pasta Recipes Perfect for Weeknight Dinner (The Stay At Home Chef) These chicken pasta recipes are perfect for busy weeknights when dinner needs to be easy, filling, and guaranteed to disappear fast. From creamy comfort food favorites to lighter pasta dinners packed with flavor, there’s something here for every kind of craving. https://thestayathomechef.com/chicken-pasta-recipes/ 2. Easy Pork Recipes Perfect for Summer Nights (The Stay At Home Chef) These easy pork recipes are perfect for summer nights when you want something flavorful without spending forever in the kitchen. From grilled favorites to slow cooker dinners and backyard classics, these recipes make warm weather cooking feel simple and satisfying. https://thestayathomechef.com/summer-pork-recipes/ 3. The Best Air Fryer Chicken Recipes (The Stay At Home Chef) Need dinner fast? These air fryer chicken recipes bring big flavor, crispy texture, and easy cleanup without turning your kitchen upside down. Perfect for busy nights, lazy weekends, and every moment in between. https://thestayathomechef.com/air-fryer-chicken-recipes/ 4. Easy Beef Casserole Recipes for Busy Weeknights (The Stay At Home Chef) These beef casseroles are packed with hearty ingredients, easy prep, and the kind of comfort food energy busy weeknights practically beg for. From cheesy classics to cozy family dinners, these recipes are made to keep dinner simple and satisfying. https://thestayathomechef.com/beef-casseroles/ 5. The Ultimate Cobbler, Crisp, and Crumble Collection (The Stay At Home Chef) If warm fruit desserts are your love language, this collection is for you. From classic cobblers to buttery crisps and crumbles, these easy homemade desserts are made for scooping straight from the dish and pretending you’re just having a tiny bite. https://thestayathomechef.com/cobblers-crisps-and-crumbles/ 6. Spinach Quinoa Lasagna Casserole (Fit Foodie Finds) A healthy and hearty option featuring spinach and quinoa layers in a lasagna-style casserole, perfect for families looking for nutritious meals. https://fitfoodiefinds.com/spinach-quinoa-lasagna-casserole/ 7. 30-Minute Pasta Primavera (Fit Foodie Finds) A quick and fresh pasta dish loaded with seasonal vegetables, ready in just 30 minutes for a light yet satisfying meal. https://fitfoodiefinds.com/30-minute-pasta-primavera/ 8. The Best Broiled Salmon (Simple Seasoning, 8 Min Broil!) (Fit Foodie Finds) A simple yet delicious salmon recipe that requires minimal seasoning and only 8 minutes in the broiler for a quick, healthy dinner. https://fitfoodiefinds.com/broiled-salmon/ ba.net/summary
|
