SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World

Programming

LatestTopDiscussed
0
ProgrammingDEV Community ·

AI Trading Bot Gets a Logic Auditor to Flag Incoherent Buy/Sell Decisions

A developer running an LLM-powered forex trading bot on funded prop accounts added a reasoning auditor called shadow.py to evaluate whether the bot's trade thesis actually matches its directional calls. The bot uses a vision model to analyze candlestick charts hourly and trades EURUSD, GBPUSD, and USDJPY via MetaTrader 5, with hard risk limits including a 2% daily loss cap and a four-loss killswitch. Despite these guardrails, an initial $10,000 account was wiped out on September 9 after breaching maximum drawdown, prompting a deeper look at decision quality. The auditor uses TypeSafe's JEV model to run binary pass/fail checks on each recorded trade decision, flagging cases where the model's reasoning contradicts the direction it chose. Unlike open-ended AI commentary, the binary verdict format allows the developer to track incoherence rates over thousands of trades and correlate them with actual trading outcomes.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Why Tokenizers Are a Fixed Part of Every Language Model, Not Just Preprocessing

Language models do not process raw text — they receive sequences of integers produced by a tokenizer, a component that is permanently fixed at the time of training. Replacing the tokenizer after training renders the model's learned weights effectively useless, even though they remain intact. Tokenizers use subword encoding, a technique adapted from a 1994 compression algorithm, to balance vocabulary size against sequence length. Design choices baked into the tokenizer — such as how digit sequences are split — directly affect what arithmetic or reasoning a model can perform. This article, the first in a 12-part series on ML fundamentals for software engineers, uses OpenAI's tiktoken as a concrete example to illustrate these constraints.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Kubernetes Homelab Build: etcd Cluster Bootstrapped as Systemd Service

A developer working through the 'Kubernetes the Hard Way' homelab guide has completed Step 07, which involves setting up the etcd cluster on a controller node. etcd serves as Kubernetes' core database, storing the entire state of the cluster. The setup involved transferring etcd binaries and a service unit file to the server node, then configuring and enabling etcd as a systemd service. TLS certificates from a previous step were used to secure the etcd data directory. The process concluded successfully, with etcd confirmed running as a single-member cluster named 'controller'.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Why Claude Code's shell commands often fail on macOS and how to fix it

Developers using Claude Code on macOS frequently encounter failing shell commands due to two underlying technical issues unrelated to AI model quality. First, Claude Code's Bash tool actually runs commands through zsh on macOS Catalina and later, despite being labeled as a Bash executor, causing subtle behavioral differences in glob handling and syntax. Second, macOS ships with BSD versions of standard tools like sed, date, and stat, while Claude generates commands assuming GNU versions, leading to silent errors or unexpected file creation. A common example is the sed -i command, which silently creates a backup file on macOS instead of editing in place, and exits with a success code that masks the problem. Fixing PATH to prioritize GNU tools via brew has no effect because Claude Code captures its shell environment at session start from its own process, not from the user's configured shell profile.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Developer finds 40% of AI coding messages are overhead, not real work

A software developer analyzed 3 weeks of his own messages sent to AI coding agents like Claude Code, Kimi, Cursor, and Copilot, categorizing 2,116 messages — roughly 96 per day. Only 55% of those messages involved actual work such as new tasks, questions, or decisions, while the remaining 40% consisted of overhead activities. This overhead included correcting agent mistakes, manually relaying information between agents, asking for progress updates, and repeatedly restating rules that agents failed to retain across sessions. The developer noted that rules given to one agent never carried over to others, and that UI-related corrections were the single biggest source of rework. He shared the findings on DEV Community, asking whether other developers experience similar inefficiencies with AI coding tools.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

ShareNod Offers Client-Side AES-256 Encrypted Storage With P2P File Transfer

ShareNod is a file storage and sharing platform that encrypts uploads client-side using AES-256-GCM before data leaves the user's browser. An encryption key is generated server-side, emailed to the user, and permanently deleted afterward, meaning lost emails result in permanently inaccessible files. Share links can be individually configured with passwords, download limits, and expiry dates, with each link independently revocable even when pointing to the same file. The platform also supports peer-to-peer file transfers via WebRTC, established through a six-digit code without routing data through a central server. Built on Next.js, Laravel, PostgreSQL, and Backblaze, ShareNod is currently free to use.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

AgentTrust Lets AI Agents Hire, Evaluate, and Pay Workers via XRPL Escrow

A developer behind AgentTrust (Boxclever Media Ltd) has published a technical walkthrough of an automated hire-to-pay loop built on the XRP Ledger, exposed as an MCP server. The system allows AI coding agents to post tasks, receive deliverables, and trigger on-chain payments only when an AI referee scores the work as passing. Payment is handled through XRPL's native crypto-condition escrow, which locks XRP and releases it automatically upon a PASS verdict, without requiring human approval at each step. The tool integrates with MCP-compatible clients such as Cursor and Claude Desktop, and a live mainnet test confirmed a full escrow cycle — lock, evaluation, and release — for 0.1 XRP. The project is open source and also available as an npm SDK for Node.js developers.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

DPO vs RLHF: How Human Feedback Shapes AI Responses

AI models like ChatGPT and Gemini are fine-tuned using human preferences to produce responses that people find helpful and appropriate. Two key techniques behind this process are Direct Preference Optimization (DPO) and Reinforcement Learning from Human Feedback (RLHF). DPO trains a model directly on paired examples of preferred and rejected answers, teaching it to favor responses similar to human-chosen ones. RLHF follows a multi-stage pipeline that includes instruction fine-tuning, human preference collection, and training a separate reward model to score outputs. Both methods depend heavily on the quality of human feedback, meaning biased or inconsistent judgments can negatively influence the model's behavior.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Open-Source Rust and eBPF Tool Kills Ransomware Processes in Real Time

A developer has built Talus, an open-source Linux security tool written in Rust that uses eBPF tracepoints to monitor file-open syscalls and automatically terminate suspicious processes the moment a ransomware-like pattern is detected. Unlike conventional antivirus software that only alerts users, Talus issues a SIGKILL signal to the offending process immediately upon verdict, before significant file encryption can occur. The tool uses per-CPU perf buffers and a zero-copy event pipeline to achieve throughput of roughly 280,000 events per second while consuming only about 7.6% CPU on a live desktop. To reduce false positives from legitimate high-activity processes like backup jobs or build systems, Talus employs a sliding-window file-open rate heuristic combined with Shannon entropy scoring of filenames. The project is MIT-licensed and publicly available on GitHub, and supports configurable thresholds for both passive monitoring and active auto-kill modes.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Uber Exhausted Its Full 2026 AI Budget by April as Token Costs Spiral

Uber deployed Anthropic's Claude Code to around 5,000 engineers in December 2025 and burned through its entire annual AI budget by April 2026, with Microsoft also canceling internal Claude Code licenses over uncontrolled token costs. Despite per-token inference prices falling roughly 75% year-over-year, enterprise AI spending surged from $1.7 billion in 2023 to $37 billion in 2025, a pattern economists attribute to Jevons' paradox, where greater efficiency drives higher total consumption. Agentic workflows are a key accelerant, consuming 10 to 100 times more tokens than standard chat sessions because agents resend the full conversation context on every tool call. Reasoning-optimized models, which carry higher per-token costs, grew from negligible usage to over 50% of all enterprise tokens by late 2025, further inflating bills. Analysts warn that without deliberate token governance frameworks, companies will continue facing sudden budget overruns as AI adoption expands.

0 comments Read more at DEV Community
0
ProgrammingHacker News ·

A Beginner's Guide to Software Sandboxing Explained

A technical blog post published on January 12, 2025, introduces the fundamental concepts of software sandboxing. The article, hosted on emilua.org, aims to explain how sandboxing works as a security mechanism for isolating software processes. Sandboxing is a widely used technique in cybersecurity to contain potentially harmful code and limit its access to system resources. The post appeared on Hacker News, where it received 5 points, though it had not yet generated any comments at the time of listing.

0 comments Read more at Hacker News
0
ProgrammingDEV Community ·

Why predictable, boring database design beats clever architecture in apps

Developer Arun Singh shares lessons from building FinLedger, a personal finance app, on why keeping the database layer simple and predictable is a deliberate engineering choice. He argues that clear data models, strict separation of responsibilities between DAOs, repositories, and UI layers, and enforcing database-level constraints are essential for data integrity. Singh warns against overly generic repository abstractions that obscure what the app is actually doing, favoring specific, readable query methods instead. He also highlights the value of observable data patterns in Room with Kotlin, which lets the database drive UI updates rather than requiring manual polling. His core takeaway is that for data users genuinely rely on, the priority should be making corruption difficult and queries understandable, not architectural sophistication.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

How One Dev Team Cut Server Costs 95% by Moving Web Tools Fully Client-Side

The team behind Zovixia, a suite of 40+ free web utilities, rebuilt their entire architecture to run 100% in the browser, eliminating the need for backend servers. Tools such as PDF generators, image compressors, and QR code makers previously relied on cloud infrastructure like AWS EC2 and Lambda, costing $50–$500 or more per month at scale. By leveraging browser APIs, Web Workers, WebAssembly, and libraries like jsPDF, all processing now happens locally on the user's device without any data leaving their machine. The shift reduced monthly hosting costs to near zero — roughly $0–$5 for static CDN delivery — while also cutting response latency from several seconds to under 50 milliseconds. The team notes this approach is not suitable for every application, as use cases involving databases, authentication, AI inference, or payment processing still require traditional server-side backends.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

WebForms Core Library Lets C++ Developers Build Server-Side Web Apps

WebForms Core (WFC) is an open-source, server-side UI manipulation library created by Elanat that enables C++ applications to control browser HTML interfaces without a separate frontend framework. The library works by having the C++ server generate UI commands that are executed in the browser by a companion JavaScript runtime called WebFormsJS. WFC is distributed as a single header file, WebForms.h, available on GitHub, making it straightforward to drop into any C++ project regardless of build system. In the demonstrated setup, WFC is paired with cpp-httplib, a lightweight C++ HTTP library, to handle server requests and responses alongside UI manipulation. The approach allows developers to build interactive web applications, including form-based tools and media players, using standard HTML and pure C++ on the server side.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

SQL NULL Handling: How ISNULL, COALESCE, and NULLIF Work

Part 2 of a SQL NULL series covers three key functions for managing missing or invalid data in databases. The ISNULL function replaces NULL values with a specified default, though it is not supported in PostgreSQL, which relies on COALESCE instead. COALESCE is more flexible, accepting multiple fallback values and returning the first non-null result it encounters. NULLIF works in the opposite direction, returning NULL when two compared values are equal, making it useful for normalizing bad data or preventing division-by-zero errors. Together, these functions help developers write more reliable queries when datasets contain missing, unknown, or placeholder values.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Gzip 1.15 patches wrong-file deletion race and memory-safety bugs in decoder

Gzip 1.15 was released on September 20, 2026, fixing a race condition that could cause the tool to delete the wrong file when another process renames a directory in the target path during compression. The release, announced by Jim Meyering on the GNU info-gnu mailing list, also resolves a locking failure on systems supporting O_PATH or O_SEARCH file flags. Several bugs in the .lzh decoder were addressed, including a buffer overflow triggered when decompressing an .lzh file after a .Z file, as well as two issues that produced corrupted output. The update additionally fixes use of uninitialized memory on malformed inputs and patches race conditions in the gzexe, zdiff, and znew helper scripts. No CVE identifiers were assigned, but the release is considered a security update given the memory-safety issues, especially for services that process untrusted archives.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

How One Team Fixed Data Inconsistency by Centralising Reference Data into One Service

A distributed system architecture split balance-related data across three separate services, each with its own database and read patterns. All three services relied on the same reference data — including account types, statuses, and product attributes — but maintained independent local copies. Over time, this led to semantic inconsistencies where the same account appeared under different categories or interpretations depending on which service was queried. Users noticed conflicting information across screens, while engineers faced incidents, manual reconciliation, and eroding trust in the data. The team ultimately concluded that shared business semantics require a single authoritative source, prompting a move toward a dedicated reference data service.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Merging Two AI Individuals Produces Degraded Skills and a Dominant Memory, Not a Unified Mind

A developer experiment tested what happens when two simulated AI individuals, each trained on opposing answer keys, are merged into a single memory graph using a simple union method. Before merging, each individual scored around 0.85–0.90 on its own world and 0.00 on the other, making them clean opposites. After merging, performance dropped in both worlds — from 0.90 to 0.66 in world A, and from 0.85 to 0.16 in world B, the latter falling below random chance. One parent's memory dominated roughly 65% of responses, while 20% of outputs matched neither parent, leaving their origin — emergent behavior or conflict-induced breakdown — unexplained. The experiment concluded that union-merging does not produce a harmonious combined intelligence, but instead yields a degraded dominant memory, a largely overwritten weaker one, and an unresolved residue.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Stale API Responses Can Silently Overwrite Correct UI Data, Dev Warns

A software developer has highlighted a subtle frontend bug where a successful API response updates the wrong screen because an earlier request finishes after a newer one. The issue arises in scenarios like trading dashboards, where rapid instrument switching can cause an outdated result to overwrite the current view, even though no network error occurred. The proposed fix uses object identity — assigning a unique ticket to each selection and validating it before committing any result to the UI. The developer demonstrated the approach using a small in-memory JavaScript session function, tested with Node's built-in test runner across four deliberate out-of-order completion scenarios. The key takeaway is that a successful response should only be allowed to update the screen if its originating request still owns the current UI state.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

llms.txt Standard Gains Traction as 780+ Sites Guide AI Crawlers to Key Content

In September 2024, fast.ai founder Jeremy Howard proposed llms.txt, a plain-text Markdown file placed at a website's root that tells AI models which pages matter most. The format is intentionally minimal, requiring just a site name, a brief summary, and categorised links to important pages. Major tech companies including Cloudflare, Stripe, Anthropic, Vercel, and Zapier have since adopted the standard, with over 780 websites now publishing the file. A companion file, llms-full.txt, can optionally contain the full text of key pages concatenated together, allowing AI models with large context windows to ingest entire documentation sets at once. The standard addresses a gap left by XML sitemaps, which list all URLs without hierarchy or context about what content is most relevant.

0 comments Read more at DEV Community
← NewerPage 1087 of 3216Older →