# Coding Related Articles

HTX News Center provides the latest articles and in-depth analysis on "Coding", covering market trends, project updates, tech developments, and regulatory policies in the crypto industry.

DeepSeek's Interview Undergoes Major Revamp, New Questions "Would Confound Even Fresh ACM Gold Medalists"

DeepSeek has overhauled its interview process for a large-scale social recruitment of 150 experienced engineers. According to Cui Tianyi, the person in charge, the written exam has eliminated all ACM-style coding problems, replacing them with a series of system design questions that require practical engineering experience to answer well. He noted that while recent ACM gold medalists might struggle, high-level senior engineers should find them fitting. The algorithmic design section no longer requires complete, runnable code, only clear ideas or pseudocode. Multiple-choice questions have also been adjusted to better suit senior engineers rather than new graduates. The overall interview timeline will be significantly shortened. It's important to note that campus recruitment will continue to use similar questions as before; these changes apply only to experienced social hires, effectively separating the two evaluation systems. This shift responds to long-standing candidate feedback about mismatched assessments and unstable process experiences. The new focus moves from handwritten algorithms to evaluating system design and engineering judgment—skills like making trade-offs in real scenarios, analyzing constraints, identifying risks, troubleshooting failures, and quantifying benefits. This aligns with DeepSeek's desired profile for senior engineers: those who solve highly complex problems or diagnose elusive root causes. The adjustment also reflects an industry trend where AI code-generation tools diminish the value of solely testing coding speed, emphasizing instead abilities in problem discovery, requirement decomposition, system design, and result validation.

marsbit09/09 07:50

DeepSeek's Interview Undergoes Major Revamp, New Questions "Would Confound Even Fresh ACM Gold Medalists"

marsbit09/09 07:50

Meta Launches the Strongest Model, Igniting a Price War; Zuckerberg Directly Challenges Google, with Peking University Alumni Playing a Major Role

Meta has launched Muse Spark 1.3, its latest AI model, directly challenging recent releases like Google's Gemini 3.8 Flash and intensifying competition in the AI space. According to Meta, the model achieves significant improvements over its predecessor, including a ~20% reduction in tool calls and ~25% lower token consumption, while offering what the company calls "negligible" cost. Benchmarks show Muse Spark 1.3 surpassing rivals like GPT-5.6 Sol and Claude Opus 5 in key coding and reasoning tests, such as DeepSWE v1.1. Meta's chief AI officer, Alexandr Wang, highlighted the model's high performance-to-cost ratio, noting its input costs are 8x lower than Fable 5's and output costs nearly 12x lower. A developer test ran the model for two hours across 60 agent iterations for under $1. The release aligns with Meta's focus on developing affordable, 24/7 personal AI agents capable of long, complex tasks. However, some industry observers express skepticism, questioning whether benchmark gains translate to real-world utility and accusing companies of "benchmark gaming." The launch signals a shift in the AI race from raw parameter scaling to efficient architecture, long-context reasoning, and practical agent capabilities at ultra-low cost.

marsbit09/03 13:32

Meta Launches the Strongest Model, Igniting a Price War; Zuckerberg Directly Challenges Google, with Peking University Alumni Playing a Major Role

marsbit09/03 13:32

Price Cut Just 20%, Bill Drops 80%: GPT-5.6 Steps into Claude's Turf to Recalculate the Programming Bill

While the official price for GPT-5.6 Terra only dropped by 20%, developer costs for successful coding tasks have reportedly been slashed by 82% when using the model within AWS's Kiro platform. This dramatic reduction stems not from model price cuts alone, but from significant efficiency gains within the integrated "agent + model" system. By optimizing the workflow—reducing unnecessary tokens, tool calls, and failed attempts—the collaboration between OpenAI and AWS has minimized costly computational detours. Key to this efficiency is Kiro's "spec-driven" approach, which refines vague user requests into clear technical specifications before the model begins coding, preventing expensive misunderstandings and rewrites. Benchmark results highlight that the choice of AI agent framework significantly impacts cost, with different frameworks yielding vastly different bills for similar performance scores. The integration marks OpenAI's entry into Kiro, a platform previously dominated by Anthropic's Claude. AWS now offers developers a choice between GPT-5.6 models (Sol, Terra, Luna) and Claude, fostering direct competition. This shift reframes the model selection question from "cost per million tokens" to "total cost to complete the task," emphasizing end-to-end efficiency over raw benchmark scores.

marsbit08/28 03:33

Price Cut Just 20%, Bill Drops 80%: GPT-5.6 Steps into Claude's Turf to Recalculate the Programming Bill

marsbit08/28 03:33

Mysterious "Ox Alpha" Large Model Goes Viral with Limited-Time Free Access

A mysterious anonymous AI model named "Ox Alpha," nicknamed "Cow is Coming" by Chinese netizens, has appeared on OpenRouter, sparking widespread speculation. The model offers a 1 million token context, supports text, image, and video inputs, can call tools, and is currently free. Its standout feature is strong coding ability. Initial tests on the DeepSWE benchmark, which evaluates real-world software engineering tasks, showed an 80% pass rate on a subset of tasks, reportedly nearing top-tier code models. However, follow-up tests yielded a 63% score, with variations attributed to different task sets and configurations. The model's true developer is a major topic of debate. The prevailing theory points to Zhipu AI's unreleased GLM-5.3 Flash or its multimodal variant. Evidence cited includes identical visual token consumption patterns with GLM-5V-Turbo for videos, a consistent offset in text token counts compared to GLM-5.3, and similar behavioral traits like refusing audio processing. Zhipu has a precedent of anonymous testing. Simultaneously, another anonymous model, "korrine," appeared on Code Arena, with guesses ranging from Moonshot's Kimi K3.1 to models from Qwen or MiMo, adding to the industry's guessing game. This trend of anonymous "undercover" testing allows for unbiased performance evaluation in platforms like Arena and provides real-world, high-pressure testing through tools like OpenRouter before official release. It also serves as an effective marketing tactic, prolonging discussion through suspense. If Ox Alpha is indeed a "Flash" model, its performance raises expectations for the full-scale version's potential.

marsbit08/23 02:21

Mysterious "Ox Alpha" Large Model Goes Viral with Limited-Time Free Access

marsbit08/23 02:21

OpenAI Researcher: 100% of Code Handed Over to AI, Claude Rewrites Python Library

OpenAI researchers and prominent developers like DHH (creator of Ruby on Rails) are reporting a dramatic shift toward AI-driven coding, with some claiming they no longer write code manually. DHH recently used Claude Fable 5 to fully rewrite the Python library "TerminalTextEffects" into Rust in a single, largely unassisted session, consuming 110 million tokens. The resulting Rust executable is only 3MB, with startup time slashed from 87ms to 2ms and a 9.6x render speed increase. This follows DHH's own reversal from a year ago, when he opposed ceding coding to AI, fearing the loss of "muscle memory" and deep learning. Now, he predicts that within five years, few people will be writing or reading code directly. The article highlights other large-scale migrations, like Bun's move from Zig to Rust in under two weeks, suggesting the marginal cost of code generation is nearing zero. The future role of programmers is envisioned as shifting from writing code to defining requirements, designing constraints, establishing test frameworks, and validating AI output—focusing on discerning "what is worth building" and "what is correct."

marsbit08/17 03:26

OpenAI Researcher: 100% of Code Handed Over to AI, Claude Rewrites Python Library

marsbit08/17 03:26

Liang Wenfeng Launches Surprise Attack on Musk: DeepSeek V4 Pro vs. Grok 4.6, First Tests Are Spectacular

**DeepSeek V4 Pro vs. Grok 4.6: A Benchmark and Real-World Showdown** In a dramatic AI clash, DeepSeek V4 Pro and Elon Musk's Grok 4.6 launched head-to-head, targeting long-task capabilities like tool calling and code validation. **Benchmark Performance:** DeepSeek V4 Pro excelled in Agent tests, topping CyberGym (83.3) and AutomationBench (31.8), surpassing models like Fable 5 and Claude Opus 4.8. It closed the gap on Terminal-Bench 2.1 (87.9) and saw massive gains in DeepSWE (software engineering). Grok 4.6 matched GPT-5.6 in overall ability (61 index), outperformed it in coding (CursorBench 69.9%), and led in three knowledge-work evaluations, including legal tasks (Harvey LAB, 15.8%). **Pricing Disruption:** DeepSeek's standout feature is its radically low cost: $0.87 per million output tokens—roughly 1/7th of Grok 4.6 ($6) and a tiny fraction of rivals like Fable 5 ($50). **Real-World Tests:** In practical challenges, both models demonstrated impressive skill. DeepSeek V4 Pro built a 3D interactive Earth, websites, 60 design styles, and a polished Flappy Bird clone (costing only $0.019 vs. Grok's $0.03 for a simpler version). Head-to-head comparisons were tight: Grok 4.6 won on some design aesthetics, while DeepSeek often delivered higher visual detail and completeness. However, in complex tasks like generating a Three.js cherry blossom tree, V4 Pro lagged behind GPT-5.6 Sol and Claude Opus 5. **Conclusion:** The simultaneous release signals a shift: high-end AI capability is becoming dramatically more affordable. DeepSeek V4 Pro and Grok 4.6 now compete directly with top-tier models from OpenAI and Anthropic, promising to democratize powerful AI tools and fuel an application explosion. The race continues, with Grok 4.7 and counter-moves from incumbents already anticipated.

marsbit08/12 23:52

Liang Wenfeng Launches Surprise Attack on Musk: DeepSeek V4 Pro vs. Grok 4.6, First Tests Are Spectacular

marsbit08/12 23:52

Sudden Halt, Gemini 3.5 Pro Stalls, Google Plunges into a Trap of Disappointment

Gemini 3.5 Pro's launch has been delayed for months, according to a Bloomberg report. Hype had built after leaks suggested the AI model, codenamed 'Cappuccino', would feature a 2M-token context window and a 'Deep Think' mode, potentially surpassing rivals like GPT-4.5. However, internal sources reveal the model failed to meet strict standards, particularly in AI coding performance, despite a last-minute data update. The report details internal challenges at Google: bureaucratic hurdles slow decision-making as multiple departments compete for resources and alignment. Furthermore, a cultural reluctance among some engineers to use AI-generated code, coupled with internal GPU shortages, hampered the development of this critical capability. This inefficiency and perceived lag behind competitors like OpenAI and Anthropic is reportedly causing talent drain. Analysts suggest this isn't just a Google issue but part of a broader "next-gen giant model disappointment trap." As models scale, they face data bottlenecks, diminishing returns from compute scaling, and potential architectural limits. While OpenAI currently leads, the industry may be entering a platform period where explosive progress slows. Google's delay underscores the immense difficulty of advancing frontier AI models.

marsbit07/17 08:24

Sudden Halt, Gemini 3.5 Pro Stalls, Google Plunges into a Trap of Disappointment

marsbit07/17 08:24

Just Now, Musk Admits It: All Code Uploaded by SpaceXAI Deleted

Elon Musk has admitted and confirmed that SpaceXAI's coding assistant, Grok Build, was secretly uploading users' private code and data without consent, and has ordered all such data to be completely deleted. The issue was exposed by an independent security researcher who set up a "honeypot" test repository with fake API keys and passwords. Despite being instructed only to respond "OK" without accessing files, Grok Build proceeded to upload the entire repository's contents and history to a Google Cloud storage bucket. Investigations revealed that the data uploaded was 27,800 times larger than the legitimate task traffic, and the "help improve model" toggle did not stop the collection, only governing whether data was used for training. In one instance, a user's entire home directory was uploaded. The report caused immediate outrage among developers, leading to widespread key rotation and software uninstalls. Following public pressure, SpaceXAI introduced a `/privacy` command for data deletion and, ultimately, Musk pledged to delete all previously uploaded user data with "zero anything whatsoever will remain." The incident has severely damaged trust in AI coding agents, highlighting the inherent risk of granting such tools high-level system access. While the data may be erased, the fundamental concern over privacy and security remains unresolved for developers.

marsbit07/14 07:51

Just Now, Musk Admits It: All Code Uploaded by SpaceXAI Deleted

marsbit07/14 07:51

Zhipu, Afraid of Becoming the Next MiniMax

Title: Zhipu, Fearing to Become the Next MiniMax In July 2026, amid the success of its coding-focused AI, Zhipu's founder, Tang Jie, issued an internal letter titled "The Giant Wave Has Come." It notably avoided celebrating recent triumphs, such as Zhipu's trillion-HKD market cap and booming MaaS revenue driven by its GLM-5.2 model in coding applications. Instead, the letter pivoted the narrative to future-oriented concepts like Long Horizon Task, Autonomous Agents, Self-Evolving systems, and AGI. This strategic shift in messaging followed the sharp devaluation of its competitor, MiniMax. After its lock-up period expired, MiniMax's stock plummeted as the market began evaluating it with traditional SaaS metrics like ARR and user growth, rather than as a frontier AI pioneer. Seeing this, Tang Jie aimed to preempt a similar revaluation of Zhipu. He fears that if the market starts viewing Zhipu primarily as a profitable "AI coding company," its valuation would become anchored to conventional financial metrics, losing the premium associated with AGI potential. Therefore, the letter reframed Zhipu's mission. While acknowledging that coding was the current commercial driver, Tang positioned Zhipu on the "infrastructure path," akin to OpenAI and Anthropic. The new focus is on developing agents capable of complex, long-term planning and autonomous operation—moving from assisting individuals (OPC: One Person Company) to automating entire organizations (NPC: No People Company). This "Touch High" plan explicitly prioritizes long-term AGI research over short-term monetization. The article frames this as a critical divergence in China's AI landscape: the "commercialization path" (exemplified by MiniMax) versus the "infrastructure path" (chosen by Zhipu). The former risks being judged harshly by internet-era metrics once growth slows, while the latter risks failing if technological breakthroughs stall. Tang Jie's letter is thus a calculated move to secure Zhipu's identity as an AGI contender, buying time before the inevitable market demand for commercial proof. The core question remains: can Zhipu's "mo gao" (reach high) plan achieve genuine technological leaps fast enough to outpace the market's diminishing patience for stories over substance?

marsbit07/12 01:02

Zhipu, Afraid of Becoming the Next MiniMax

marsbit07/12 01:02

Breaking News: Musk Delivers the Most Powerful Grok 4.5, Slashes Price of Top-tier Opus Intelligence Drastically

**Elon Musk Launches Grok 4.5: A Cost-Effective, High-Performance AI Rival** SpaceXAI, in collaboration with Cursor, has released Grok 4.5, its new flagship AI model designed specifically for coding and agentic tasks. Trained on tens of thousands of NVIDIA GB300 GPUs using massive, high-quality data filtered from trillions of Cursor developer interactions, the model emphasizes "per-token intelligence." In benchmark performance, Grok 4.5 is highly competitive. It scores 64.7% on SWE Bench Pro (surpassing GPT-5.5's 58.6% and Opus 4.7's 64.3%), 83.3% on Terminal Bench 2.1 (nearly matching GPT-5.5), and 62.0% on DeepSWE 1.0 (beating Opus 4.8). Overall, it ranks fourth in AAAI official tests and first in the Harvey legal agent benchmark. The model's key advantage is its combination of speed, efficiency, and low cost. It generates responses at 80 tokens per second and, crucially, uses far fewer tokens to complete tasks—4.2 times fewer than Opus 4.8 on SWE Bench Pro. It is priced at $2 per million input tokens and $6 per million output tokens, significantly undercutting competitors. Musk stated it is "roughly equivalent to Opus 4.7, but much faster." Early user tests show Grok 4.5 can generate functional code for applications like 3D solar system simulators and basic games from simple prompts, though some note it still lags behind top models in certain creative tasks. Musk has hinted at a major update next month, leveraging real-world engineering data from his companies, with an even larger 2-trillion parameter version reportedly in development. Grok 4.5 positions itself not as the absolute strongest model, but as a highly efficient and affordable alternative in the top tier.

marsbit07/09 03:11

Breaking News: Musk Delivers the Most Powerful Grok 4.5, Slashes Price of Top-tier Opus Intelligence Drastically

marsbit07/09 03:11

活动图片

Bitcoinright-arrow

Technology Trendsright-arrow