How AI Helps

Mistral launches Large 4 in API preview, leading a test of finding and fixing security bugs at 81.7%

Artificial Analysis used 131 tasks from C/C++ projects such as FFmpeg and CPython. A pass meant showing a crash, supplying a patch that stopped it, and keeping existing tests passing.

The score can include fixing a different real bug from the one the task targeted.

Mistral plans to release the model for download by the end of October. That could let teams test this repair workflow on their own servers.

Codex can keep working without another “keep going”

An AI assistant investigates a bug, finds a clue, then stops with a polite progress report. The bug is still there. Apparently, your job is now to press the imaginary “please continue” button.

Goal mode in Codex addresses that pause. Codex, OpenAI’s coding agent, stores an objective in the conversation. After a turn ends, it can check whether the objective is met and automatically continue if work remains and the budget allows.

That matters when the next move depends on the last discovery. A slow app might need an investigation, a code change, then a fresh speed test. If the change barely helps, Codex can use that result to choose its next attempt.

The useful detail is a finish line it can check. “Faster” is vague. A defined speed target, with existing tests still passing, gives it evidence of success. That requires access to the relevant project and tests.

Users can pause or clear the goal. A blocker or budget limit can also stop progress.

The human’s job shifts toward defining what “finished” means. The imaginary button gets a quieter afternoon.

Narrow down a bug’s cause with a Claude Code team

When a bug has several plausible causes, Claude Code’s experimental agent teams can investigate them separately and challenge each other’s evidence. Each teammate uses extra tokens, so this suits a stubborn bug more than a small fix.

You need a local code project, steps that reproduce the bug, and relevant logs or a failing test. These instructions use a macOS or Linux terminal with Claude Code installed and signed in, through a supported subscription or funded API access.

1. Open your terminal in the project folder and start a session with teams enabled:

CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1 claude --teammate-mode in-process

All teammates will work in the same terminal; no extra terminal software is needed.

2. Paste this prompt, replacing the bracketed fields:

Investigate this bug without editing files.

Bug: [actual behavior and expected behavior]
Reproduce it: [steps or test command]
Evidence: [log excerpts or local file paths]

Create an agent team with three teammates. Assign each a different plausible cause based on this evidence. Give every teammate the full bug details.

Ask them to inspect the relevant code and run targeted tests where possible. Have them message each other with evidence that challenges the other explanations.

Wait for all three. Report the best-supported cause, file and line references, theories ruled out and why, and anything still uncertain. Propose the smallest test that could confirm or reject the leading explanation. Do not apply a fix yet.

3. Check the evidence before accepting the diagnosis. Run the proposed test and compare its actual output with the team’s prediction. If it disagrees, paste the result back and ask the team to revise its explanation. Agreement between agents alone does not confirm a cause.

4. When finished, ask the lead to shut down all teammates:

Ask all teammates to shut down. Give me a final summary of the evidence and the next debugging step.

HackerRank's Chakra AI interviewer leaves beta

HackerRank says Chakra has interviewed over 500,000 job candidates. Their average rating of the experience is 4.8 out of 5. This shows they liked the interview, but does not prove it helps companies hire better developers.

During the interview, candidates use AI to fix bugs in an app. Chakra asks them to explain their choices.

One example on HackerRank's website shows a candidate whose code looked reasonable. Yet they could not explain the changes or why they accepted the AI's suggestions.

Hiring teams get scores with evidence from the candidate's code and a written record of the interview. People decide who to hire.

For developers, writing code with AI is only part of the test. They also need to understand the code and explain their decisions.

Turn a pixel image into an editable vector

Recraft’s AI vectorizer converts an existing raster image, such as a PNG, into an editable SVG. It can help turn a small logo or simple illustration into artwork that stays sharp when enlarged.

You upload an image, select “Vectorize” and export the result. Recraft Studio also lets you recolour the SVG or reduce its colour count to simplify the design. The converter is available through an API for automated workflows.

Recraft describes it as a specialised AI model, but the public sources reviewed do not explain its training data or architecture, or provide tests measuring how accurately it preserves the original artwork.

AI may reach the person everyone has stopped arguing with

At a family dinner, someone announces that the Moon landing was staged. Again. Everyone knows their lines, and the potatoes are getting cold.

These arguments can make a person seem impossible to reach. A study suggests there may still be room for movement.

In a 2024 study by Costello, Pennycook and Rand, participants described a conspiracy theory they believed and the evidence they found convincing. An AI then discussed those particular claims with them. It had been instructed to argue against the belief.

After the conversations, belief strength fell by about 20% on average. The reduction persisted for at least two months. That means people became less convinced; it does not mean everyone changed their mind.

The intriguing possibility is that some people we consider unreachable may never have had their particular evidence patiently answered. A general explanation can miss the photograph, detail or apparent contradiction that keeps someone convinced. AI could make that individual attention more widely available, without another relative sacrificing their dinner.

This was a controlled study, so it cannot promise peace at the next family gathering. It does suggest that giving up on an argument need not mean the other person is beyond persuasion.

That possibility also makes the chatbot’s assigned goal matter. Here, it was directed to challenge conspiracy beliefs. A convincing conversation is not itself proof that its conclusion is true. The hopeful result is that people can reconsider; the responsibility lies in what we try to convince them of.

Grok can turn a conversation into a new Tesla destination

A trip starts with a plan. Then someone gets hungry, coffee becomes urgent, and the plan enters negotiations.

In compatible Teslas, Grok can help with the map side of that discussion. Its Assistant personality supports spoken requests to find, add and change navigation destinations. A conversation about a coffee stop can lead to that stop appearing in the car’s navigation.

That connection is what makes this useful. Grok handles the conversation; Tesla’s navigation system handles the route. The driver can discuss where to go and have the resulting choice passed straight to the map, without entering the destination separately.

There are practical limits. The beta requires an AMD infotainment processor and supported vehicle software. It also needs Wi-Fi or Tesla’s paid Premium Connectivity, plus precise-location sharing for location features. Availability varies by vehicle and market. Tesla’s support page says no Grok subscription is needed to start, though usage limits apply and continued use may require signing in.

The interesting shift is small but concrete: a chatbot’s answer can become an action inside something people already use. In this case, the family debate about coffee can at least produce a destination.

Get a small bug fix from a GitHub issue

Claude Code GitHub Actions can work on a bug while your laptop is off. A useful first task is a reproducible error with a clear expected result.

You need a repository on GitHub.com with Actions enabled, admin access for setup, and Claude Code and GitHub CLI installed locally. Authentication uses a Claude subscription token or an API key; GitHub Actions usage is accounted for separately.

1. Connect the repository once.

In its local checkout, run gh auth login, then start claude. Inside Claude Code, run /install-github-app.

Follow the prompts to install the app and configure authentication. Continue with Actions setup and select the workflow that responds to @claude mentions. Create and merge the workflow pull request opened by the installer. Official setup guide.

2. Open an issue describing one bug.

Include the steps to reproduce it, the actual result and the expected result. Add relevant file names and the test command if you know them.

For example, if an empty search crashes your app, describe exactly which page and action cause it. Then post this comment from an account with repository write access:

@claude Fix the empty-search crash described above. An empty query should show “Enter a search term” without sending a search request. Keep normal searches working and avoid unrelated changes.

Add a regression test using the existing test framework. Report which checks you ran and which you could not run. Prepare the fix on a new branch for review.

3. Turn the result into a reviewed pull request.

Claude updates its issue comment with progress. For issue requests, it creates a branch and can return a link to a prefilled pull request page. Follow that link and create the pull request. Documented behavior.

Review the changed files. Run the regression test and try both an empty search and a normal search. If Claude could not run checks, run them yourself before merging.

Strata update lets Codex CLI use a 125-billion-parameter AI model running on a home PC

Strata is software that runs AI models locally. It uses a compressed model and draws on the graphics card, CPU, system memory and storage. It supports Windows and Linux.

The update adds an interface that lets Codex CLI send requests to Strata. Qwen3.8-Flash-Next supplies the model responses, while Codex handles tool calls and returns their results to the model. In the v0.1.39 release notes, the maintainer reports testing this loop.

The reference PC has an RTX 5070 with 12 GB of video memory and 64 GB of RAM. The model does not fit entirely in the graphics card’s memory.

Developers can use a familiar coding interface while keeping model requests on their own machine. Tools can still make network calls, so a local model does not make every task offline.

Voice control that survives an internet outage

The internet is down. Your video call has frozen, but the desk lamp still responds to your voice. Apparently, switching on a light never needed a trip to a distant data centre.

Home Assistant’s local voice setup makes this possible by keeping the conversation inside your home. Speech-to-Phrase recognises supported commands, Home Assistant handles the request, and Piper speaks the reply. Even a small computer such as a Raspberry Pi 4 can run these speech tools, without sending recordings to an outside speech service.

The crucial detail is the path from your voice to the bulb. The lamp must also support local control. A bulb that depends on its manufacturer’s cloud still depends on the internet, however clever the voice recognition beside it becomes. Your home network and the computer running Home Assistant must stay powered and working.

There is a tradeoff: setup takes some effort, and Speech-to-Phrase understands a limited set of commands rather than unrestricted conversation.

For a desk lamp, that may be plenty. It needs to understand the request, not discuss your reading choices. Local AI independence can be as ordinary as a light that still listens when the internet stops.

Copilot can answer questions about your in-person meetings

Who promised to send the revised proposal? Microsoft Copilot’s Record feature lets you ask after the meeting ends—even if it never happened in Teams.

The mobile app records the conversation, saves the audio to your work OneDrive, and creates a transcript and summary in Copilot Chat. You can ask follow-up questions, identify commitments or draft a recap. The original audio stays available.

Each recording can last up to 120 minutes. Recording continues while you use other apps on your phone, so you can check a document or your calendar without stopping it.

Gemini can control several smart home devices in one request

Switching off the bedroom lights and TV can be one request to Gemini on Android. No need to open separate controls for each device.

Its Google Home connection works with compatible lights, thermostats and other smart home devices. You can name the devices or describe the change you want in a room. Google Home connects the devices; Gemini interprets your request.

For example, you can ask it to adjust the air conditioning for sleep or start a robot vacuum in the kitchen. More precise controls work too: dimming lights to 50%, lowering the temperature by two degrees or pausing the TV.

To connect it, open Gemini with the account you use for Google Home. Ask it to control a device and follow the connection prompt. If Gemini doesn’t pick Google Home, include @Google Home in your request:

@Google Home turn off the bedroom lights and TV

Once connected, you can also start Google Home routines by voice or text.

Nurses say an AI scheduling tool creates extra work

Timpani is an AI tool built by Palantir for HCA Healthcare, a hospital operator. It helps create nurses’ work schedules: who works on which days and shifts.

WIRED interviewed six nurses about the tool. They said they had to spend time fixing their schedules. One nurse, Amber Retzloff, said it failed to follow more than half of her 50 shift requests. Nurses also reported shifts with too few nurses or too few experienced staff.

HCA says Timpani saves managers time. It also says nursing leaders still have the final say on schedules.

To understand whether the tool helps overall, we need to count both the time managers save and the time nurses spend fixing schedules.

Perplexity releases an AI model for routine business decisions

The new model helps developers build apps that sort incoming information into predefined categories.

For example, a customer reports a broken integration. The model can assess which support team should handle it and how urgent it is. It returns probabilities for the available options, so an app could route the request automatically or send an uncertain case to a person.

This could reduce manual sorting in customer support. The model handles classification; it does not write replies to customers.

Developers can use Perplexity’s paid API or download the model to run on their own hardware. Running it locally requires a powerful GPU.

Create a reusable bug review for your Git branch

Before merging code, you may want separate checks for logic errors and missing tests. Grok Build can save that process as a workflow, with another agent checking the findings before you read the report.

You need a local Git repository, a branch with committed changes, and Grok Build installed and signed in. Tests require the project’s dependencies.

1. Follow the Grok Build setup guide. In your terminal, open the project folder and run grok. First launch opens a browser for authentication.

2. Paste this into Grok Build:

/create-workflow Create a project workflow named review-changes. Accept a target argument containing a Git diff range. Review only changes in that range, reading surrounding code when needed. Use two parallel reviewers: one for logic bugs, one for missing tests and edge cases. Then use a separate verifier to check each finding against the code and available tests. Do not edit project files or install dependencies during a review. Return one report with confirmed findings, file and line references, likely impact, and reproduction steps. Put unverified concerns in a separate section and say which tests could not run.

Answer its questions about scope. Ask it to save the workflow in the project. Grok creates and smoke-checks the workflow; that check does not prove its reviews will be correct.

3. Launch it with the comparison you need:

/workflow review-changes {"target":"origin/main...HEAD"}

This example reviews committed branch changes since the shared ancestor with origin/main. Replace that reference if your base branch has another name, and fetch the remote first if your local reference is stale.

Use /workflows to view progress. For a later branch, run the saved workflow again with the appropriate diff range. The workflow documentation covers these commands.

Before acting on a finding, open the cited lines and try its reproduction steps or relevant test. A useful finding should demonstrate an actual failure; a report with no findings is not proof that the branch is bug-free.

A half-remembered tune can be a search query

You remember a song from childhood car journeys. The chorus is still there, but the title and every useful lyric have vanished. Your brain kept the soundtrack and threw away the label.

Google’s Hum to Search can work with that fragment. It compares a hummed melody with recorded songs, even when your voice, pitch or speed differs from the original.

The clever part is making two very different sounds comparable: your kitchen humming and a studio recording with a whole band. The model turns each into a numerical representation that emphasizes the melody. Similar melodies should end up close together, despite the missing drums or your unexpected solo career. That is the principle behind Google’s system.

The results are possible matches, with listening providing the moment of recognition. A familiar phrase might bring back the whole chorus. A tune invented by someone in your family, however, may have no recording in the catalogue to match.

What makes this useful goes beyond music recognition: a memory can become searchable before you can put it into words. Sometimes you have no idea what to type, yet still know exactly how something goes.

Perplexity brings paid market data into an AI conversation

Perplexity’s Premium Sources lets its Computer assistant search selected data from PitchBook, CB Insights and Statista: private company funding, industry research and market estimates.

Computer picks relevant sources and cites the originals. Small teams can explore competitors’ funding and market demand in one conversation, without separate subscriptions to those providers.

Access covers selected material, not full databases. It requires Pro or Max on the web and available Computer credits. Runs consume credits; Pro has no recurring monthly allowance, while Max includes one.

Google unveils Gemini 4 Argon, reporting a 2.7× speedup in video decoding

Google says its agents replaced 32,000 lines of code in an existing Rust version of the libgav1 decoder, preserving identical video output.

They repeatedly tested speed and inspected the compiler’s output. Their safe Rust code let the compiler generate parallel instructions automatically.

The gain is measured against the earlier Rust version. Google says it brings performance closer to optimized C++.

Model access is initially limited to trusted cyber defenders through Fairwind, with no firm date for a wider release.

DeepMind adds hidden AI labels to protein designs

A watermark can show that an image was made with AI. DeepMind’s SynthID Bio brings that idea to proteins by embedding a detectable pattern in the order of their building blocks, called amino acids.

In a Nature study, researchers added these marks to existing protein designs. Lab tests showed that the marked proteins still attached to their intended targets about as well as unmarked versions.

The approach could help labs that make DNA check incoming orders for AI designs. It could also help researchers identify AI entries in biological databases.

This is an early demonstration. Changing a protein’s sequence can largely erase the mark, so it cannot reliably identify every AI-designed protein.

Bad bass? The chair may matter more than new speakers

Bass booms at the desk and almost disappears a few steps away. Reflections from room surfaces can reinforce or partly cancel bass frequencies. Speaker placement and where you sit both matter.

Room EQ Wizard (REW) measurements taken with a calibrated microphone show peaks and dips at your listening position. An AI assistant can use those graphs and a room sketch to explain possible causes and suggest a small speaker or chair move to test.

REW’s room simulator can help explore positions, though its rectangular room model is an approximation. Fresh measurements and listening tell you whether a move helped.

AI turns a confusing graph into a practical experiment with equipment you already own. The next audio upgrade might simply be moving your chair.