SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World

Programming

LatestTopDiscussed
0
ProgrammingDEV Community ·

From Docker Compose to Terraform on AWS: Why Running Locally Is Not Enough Evidence

A developer explains that a docker compose setup proves code runs but not that a system is secure or auditable, which matters for patient-identifiable NHS data. The Terraform configuration describes a 46-resource AWS stack and is planned in CI using mock credentials, since an earlier real deployment was torn down. The author also found that the repository's Gitleaks configuration disabled all default rules, so the secret scan had been silently ineffective for months. Porting the GitLab pipeline to GitHub Actions exposed several such problems once it ran for the first time. The piece argues that continuous checking of infrastructure code is needed to show a system works, not just to run it.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Failed AI Visibility Checks Should Not Be Counted as Brand Misses

A reporting guide argues that dashboards tracking brand mentions in AI answers should not treat failed or pending checks as misses. Using an eight-check example with five completed, two failed, and one pending, it notes that dividing by all rows yields a misleading 25%, while dividing by classified completed answers gives 40%. The author recommends showing coverage alongside the rate, distinguishing unclassified answers from negative ones, and returning no rate when nothing is classified. The accompanying code requires a strict boolean for classification and was tested against boundary cases. The guide also warns that retries counted as separate observations can distort the denominator upstream of the reporting logic.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Data platform design, not just the model, determines the risk of putting AI agents on business data

The article argues that much of the risk in deploying AI agents on organisational data lies in the data platform rather than the model itself. It uses billing scenarios such as mid-period plan upgrades and downgrades to show how commercial rule decisions shape the revenue figures and invoices that analysts and customers depend on. Agents, the author says, need documented rule history, attributed corrections captured at the time of entry, and enforced limits on what they can change. An illustrative support case shows an agent misreading stale price data, issuing credits that bypass per-transaction approval limits, and acting on unscreened instructions, with costs traced mainly to platform design. The piece also stresses that business definitions should be owned by the business, with analytics and data engineers handling implementation and guarantees.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Two PHP Flaws Rated 9.2 Under CVSS v4.0 to Be Fixed in Upcoming Release

Two vulnerabilities in the PHP codebase have each been rated 9.2 on the CVSS v4.0 severity scale, indicating a serious risk. The flaws reportedly arise from coding oversights, gaps in security testing, and previously unknown weaknesses in the language. If exploited, they could allow attackers to run arbitrary code, gain unauthorized access, and compromise entire web applications. A forthcoming PHP security release is slated to patch these issues, and administrators are being urged to apply it promptly.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

FACTS.md pattern keeps Claude from inventing prices in customer replies

Small businesses using LLMs to answer customers often get confident but wrong replies, such as unapproved discounts or unlisted delivery promises. The proposed fix is to have the model answer only from a single FACTS.md file that lists prices, hours, and policies. If a fact is absent, the model replies with a holding message and flags the gap to the owner as a to-do. The author tested this on a fictional coffee shop across 11 prompts, which surfaced issues like replies in the wrong language that were then added as rules. The approach is only as reliable as the FACTS.md file, and the author recommends a human review drafts before sending.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Three ways to take a full-page screenshot in Chrome, from built-in tools to one click

Chrome can capture an entire web page, top to bottom, without installing any software, though the feature is not prominently displayed. The article describes using DevTools' Command Menu to choose "Capture full size screenshot," which saves a PNG to the Downloads folder. It also explains capturing a full page at a specific mobile width through DevTools' Device Toolbar, with a warning about device pixel ratio affecting image size. Finally, it presents a free Chrome extension called Capture by RATdeck, disclosed as made by the author's team, which automates scrolling, handles sticky headers and scrollable panels, and includes built-in editing tools.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Collapsed sidebar clipped a dropdown menu due to implicit overflow-x behavior

A QA report flagged that the '+ Create' dropdown was cut off at the right edge when the sidebar was collapsed to 60px, while it rendered correctly at 244px. The cause was that setting overflow-y: auto forces an unspecified overflow-x to compute to auto, turning the sidebar into a scroll container that clipped the 190px menu. The developer could not remove the vertical scroll, so the menu is now positioned with position: fixed only in the collapsed state, using coordinates from the button's bounding rectangle. Console checks confirmed the menu sat outside the sidebar and fully within the viewport. The fix was limited to the collapsed state to avoid disturbing existing expanded-state behavior.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Kinewright adds Wheels, a kinematic engine for four-wheeled bodies

Kinewright, a project with an existing human engine, now includes a second engine called Wheels that models how a four-wheeled body moves. It is kinematics only, with no tyres, suspension, or physics, and all reported figures are traced from the engine's own runs rather than measured. The engine has passed three stages, covering a baseline turning-radius check, a capability contract across shell types and nine manoeuvres, and a driveable shell that a person can control from a keyboard. Each drive is recorded in a journal that replays identically across Linux, Windows, and macOS. The project explicitly is not a path planner, autonomy stack, vehicle simulator, or hardware controller, and it still lacks obstacles and collision handling.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

hivemux: a tmux sidebar that flags which Claude Code session needs input

Developer Esteban Kim built hivemux, a free, MIT-licensed tmux sidebar for managing several Claude Code sessions at once. The sidebar shows whether each session is running, waiting for user input, or finished, and sends a macOS notification when one needs attention. Keyboard shortcuts and clicks let users jump straight to a session, including the one that has waited longest. Because state lives in tmux, sessions survive closing the terminal, and tmux-resurrect restores them with claude --resume after a reboot. The tool runs without a server or telemetry and currently targets macOS, with Linux likely but untested.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Claude and a weather MCP server used to plan a Menorca trip

A Meteosource employee planned a family trip to Menorca by connecting Claude to the company's weather MCP server using only a free API key. The model requested forecasts for two locations on opposite sides of the island, then hourly data for the specific south-coast village where the group was staying. Using that data, Claude produced a day-by-day itinerary matched to rain, temperature, and wind conditions. It also linked wind direction to geography, sending the group to sheltered south-coast beaches on a windy north-wind day. Follow-up requests, such as adding lighthouse visits, were answered with suggestions split by which coasts were suitable under the forecast conditions.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

US Army solicits self-erasing autonomous offensive agents, raising forensic concerns

The US Army has opened a solicitation for ECHOTRIBBLE, a framework of offensive autonomous agents that reportedly infiltrate systems, replicate to other machines, neutralize targets, and delete themselves before detection. The named targets include iPad, macOS, and Windows devices, and the design calls for minimal human oversight. Analysts argue that the self-destruct feature threatens forensics, since traces left on endpoints would be erased and attribution would become difficult. The same week, Microsoft expanded its Guardian Agents offering, which emphasizes identity, logging, and containment for AI agents. The article also notes that legal responsibility for breaches caused by autonomous agents remains unassigned.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Where Coding Agents' Token Costs Come From, and Seven Ways to Reduce Them

A coding agent's bill can rise even for a simple question, because each request re-sends the entire session history. Claude Code includes the full conversation with every request, and tool use adds further requests that resend that context. Prompt caching lowers the price of repeated context but does not eliminate it, while long sessions, cache misses, background tasks and compaction all add cost. Output tokens, including extended thinking, are priced at five times the base input rate, so the model's responses are a significant driver too. The article recommends measuring usage first, then applying seven levers in order: caching, scoping, controlling output, routing, batching and budgeting.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Brief: A Browser-Only AI Tool That Distills Notes in Under a Second

A developer has released Brief, a free web app that converts raw text and Markdown files into keypoints, study Q&A, executive summaries, or action items. The app has no backend database, sending requests directly from the browser to Groq Cloud using the user's own API key so that data never passes through an intermediate server. Its creator says Groq's LPU hardware lets even lengthy inputs, such as a 4,000-word lecture transcript, render in under a second. Brief supports drag-and-drop uploads of .txt and .md files and runs on Firebase Hosting without requiring account creation. The developer is seeking feedback and feature suggestions on which output formats would be most useful.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Quantum Machine Learning Could Reshape AI by Moving Beyond Classical Hardware Limits

Artificial intelligence has advanced largely on classical silicon hardware such as GPUs and TPUs, but these chips face physical limits in transistor scaling and energy use. The article presents Quantum Machine Learning as a possible successor, combining quantum mechanics with neural network design to handle complex, high-dimensional data. It explains that qubits exploit superposition and entanglement, which could let quantum neural networks represent certain probability spaces more naturally than classical models. Potential uses cited include molecular modeling, financial risk optimization, and logistics optimization. The author expects near-term systems to be hybrid, with classical GPUs handling preprocessing while quantum processors accessed via cloud APIs tackle optimization tasks.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

How Browsers Convert CSS Logical Properties into Physical Styles

Developers often find that CSS logical properties like margin-inline-start do not always behave as expected when text direction changes. Traditional physical properties such as margin-left stay fixed to the box edges, while logical properties are defined relative to the writing mode and text flow. Browsers translate logical properties into physical ones during style computation, mapping inline-start to left in LTR horizontal text and to right in RTL. In vertical writing modes, the inline and block axes swap, so logical properties can map to top or bottom edges instead. The layout engine ultimately only works with physical properties, which explains many unexpected layout results.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Three Policy Gaps That Let Session Keys Exceed Their Permissions

Session keys and agent wallets are meant to act within tightly scoped limits, but their policy code often contains flaws that undermine those limits. One common gap is that spending policies count only transfers, so an approval call can grant an attacker a lasting claim on funds that survives revocation of the session. Allowlists keyed on contract addresses are another weak point, since allowing a contract exposes every function on it, including dangerous generic execution or admin methods. Fixed-window spending caps can also be exploited at window boundaries, allowing roughly double the intended amount in a short span, a risk amplified by automated agents that retry rapidly.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Forma Chat: open-source team chat with auto-drafted standups and bring-your-own AI

Forma Chat is an Apache-2.0 open-source team chat that runs on your own server as a single Go binary backed by SQLite. It offers channels, direct messages, threads, search, and task creation from messages, along with daily standup drafts generated from each person's task events without using any AI model. Drafts stay private to their author until submitted, and private work is never summarized automatically. Optional AI agents can use a team's own API key or connect a user's Claude Code over MCP, and answers are limited to what the asking channel can already see. The project is in early release, with files, huddles, and mobile apps not yet available.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Sorting Prompts by Topic Speeds Up Mixture-of-Experts Batch Jobs

A mixture-of-experts model must read every expert that any request in a batch selects, so mixed-topic batches force many expensive weight reads. Because similar prompts tend to route to the same experts, grouping them together reduces the number of experts read per step. Hugging Face measured this on unchanged vLLM and found throughput gains of 5 to 17 percent. The approach works because a batch server fills its slots from the front of the queue, so neighboring prompts decode side by side. The technique matters chiefly for offline jobs on self-hosted mixture-of-experts models, and clustering prompt embeddings can achieve similar gains when labels are unavailable.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Neuro-Symbolic AI Pairs Deep Learning With Formal Logic to Address Trust and Explainability

Deep learning systems have achieved major advances in tasks like image recognition and text generation, but they operate as black boxes whose decisions cannot be verified or explained. They also tend to produce confident false statements when they fail. Neuro-Symbolic AI responds to these weaknesses by combining neural networks, which process unstructured data, with symbolic logic engines that apply explicit rules and auditable reasoning. Proponents argue this hybrid design is better suited to safety-critical areas such as autonomous driving, healthcare, and financial compliance. The approach is presented as a step toward AI that can both recognize patterns and understand the rules governing them.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Disconnecting a Crypto Wallet Does Not Remove Smart Contract Token Approvals

Many crypto users assume that ending a website session disconnects their wallet's permissions, but the two are separate. Token approvals recorded on-chain under ERC-20 standards can stay active after a user leaves a decentralized application. Some apps request allowances far larger than a single trade needs, and those leftover permissions may let a spender move tokens up to the remaining limit. Such allowances can become a risk if the spender is compromised or malicious, or if a phishing site disguised the approval as routine. Users are advised to review active allowances with a reputable checker and revoke any they no longer need.

0 comments Read more at DEV Community
← NewerPage 191 of 2982Older →