Bilingual

Nano Banana 2 Lite
23 days ago
- Nano Banana 2 Lite is a high-speed, cost-efficient Gemini Image model designed for fast image generation and editing.
- It enables rapid prototyping and iteration, allowing creators to generate thousands of images at a low cost compared to heavier models.
- The model maintains control and accuracy, supporting character consistency, precise visual editing, and real-world knowledge integration.
- Example applications include Space Lift for interior design, Gridscape for interactive learning, Peek-A-Word for educational visuals, and Anywhere for travel postcards.
- Performance benchmarks show strong results in image editing and generation, with low latency per 1k resolution image.
- Testimonials from companies like Figma Weave, Manus AI, Artlist, Weekend, and Latitude highlight its speed, reliability, and quality for creative and real-time workflows.
- Users are advised to provide detailed prompts for better results and to verify outputs, as the model may struggle with fine details, text accuracy, and complex edits.
- Safety features include content filtering and SynthID for watermarking AI-generated images, with ongoing improvements in safety and consistency.
- Legal disclaimers note that LLMs like Gemini 3.1 Flash-Lite Image may produce inaccurate or offensive content and should not be used for professional advice.
GLM5.2 on AMD MI355X at 2626 tok/s/node at over 2x lower cost than Blackwell
20 days ago
- AMD MI355X GPUs offer significantly lower cost per GPU (about 2.75x cheaper) compared to NVIDIA Blackwell B300, providing a cost-effective solution for AI inference.
- Wafer achieved 2626 tok/s/node aggregate throughput with GLM5.2 on AMD MI355X at 2.4 RPS, achieving 80% of B200 performance at over 2x lower cost, and 213 tok/s in single-stream decoding.
- Optimizations included quantizing GLM-5.2 to lossless MXFP4 with AMD Quark, using sglang as the inference framework, fixing speculative decode issues, and tuning MoE kernels to improve prefill performance.
- AMD faces challenges with day-0 support and software friction for frontier models, but the performance gap is narrowing as kernel and model optimizations improve, reducing reliance on custom kernels.
- The study highlights that achieving state-of-the-art performance on AMD is increasingly about support and optimization rather than inherent software limitations, indicating erosion of NVIDIA's CUDA advantage.
A verification loop 4x'd DeepSeek's intelligence, matching Opus at 1/7 the cost
16 days ago
- Verification loops significantly improve coding agents' ability to complete multi-step web development tasks by catching bugs early.
- Using IronBee (a verification layer) with DeepSeek (low-cost model) increased task completion from an average of 6.8 to 17 out of 20 tasks, matching the performance of Claude Opus (frontier model).
- DeepSeek with IronBee achieved similar results to Opus alone but at about 1/7th the cost ($2.4 vs. $15 per run), showing verification can reduce reliance on expensive models.
- The study emphasizes that verification helps most with bugs that only appear when the running application is tested, especially in sequential tasks where early errors compound.
- Limitations include using only one project and model pair initially, with plans to expand testing to more models, projects, and a cost-matched ablation study for broader insights.
Distributed System Is Slower Than a Laptop
16 days ago
- A 2015 paper showed single-threaded laptop outperformed distributed systems on large-scale graph processing tasks.
- Many distributed systems fail to beat a single competent thread even with hundreds of cores, leading to high COST (Configuration that Outperforms a Single Thread).
- Companies pay millions for distributed systems when simpler, cheaper single-machine solutions suffice and perform better.
- Hardware improvements have made single servers extremely powerful, but distributed-by-default reflexes persist without justification.
- Distributed systems often introduce unnecessary complexity, higher costs (compute, engineering, incidents), and slower performance.
- Frameworks, career incentives, vendor pricing, and misconceptions about scaling drive unnecessary distribution.
- Legitimate reasons for distribution include data size, availability needs, geographic latency, and organizational partitioning.
- Building a single-machine baseline before adopting distributed architectures can save costs and improve design decisions.
Performing Live Migrations of VMs
9 days ago
- Sailboxes offer a cost-effective pricing model, charging only for active resource usage per second, making long-horizon agent workflows scalable.
- Live migration technology enables Sailboxes to move VMs between hosts seamlessly, with minimal downtime, supporting ephemeral compute and load balancing.
- A network proxy handles open connections during migrations, spooling packets to maintain connectivity and providing additional features like whitelists and credential injection.
- P2P migrations use userfaultfd for efficient memory transfer, paging in data on demand and using larger memory chunks to reduce latency.
- Elastic memory consumption via virtio-mem allows Sailboxes to dynamically adjust memory usage, optimizing resource allocation and speeding up migrations.
- Sailboxes charge $0.015/vCPU-hour, significantly cheaper than competitors, with billing based solely on active usage for cost-effectiveness.
- Trade-offs include occasional latency spikes during migrations, but these are infrequent and manageable for long-horizon agents.
- Sailboxes are now generally available, with ongoing infrastructure improvements planned for fleet utilization and caching systems.
Everyone says Chinese AI is dirt cheap
8 hours ago
- Kimi K3 is not the cheap option; it costs similarly to Western flagships and performs mid-table on cost per test.
- Chinese models dominate as the 'doer' (coding execution), being 19 times cheaper than Western equivalents for the same results.
- Three traps include hidden billing for reasoning tokens, models failing in real tooling despite leaderboard rankings, and unstable routing on cheap aggregators.
- The optimal setup uses a frontier planner (e.g., Grok-4.5 or GPT-5.6 luna) paired with a cheap Chinese doer (e.g., DeepSeek flash).
- Cost should be measured per verified test pass, not per token.

Hasty Briefsbeta

#cost efficiency

Nano Banana 2 Lite

GLM5.2 on AMD MI355X at 2626 tok/s/node at over 2x lower cost than Blackwell

A verification loop 4x'd DeepSeek's intelligence, matching Opus at 1/7 the cost

Distributed System Is Slower Than a Laptop

Performing Live Migrations of VMs

Everyone says Chinese AI is dirt cheap