What's New

AI developments from the last 24 hours

Google's Cheaper Gemini Update Cuts Coding Costs, Boosts Automation

Google released Gemini 3.7 Flash, its latest budget-tier model for coding and automation tasks, just three weeks after Gemini 3.6 Flash. Google says the new version posts sizable gains on coding and web-development benchmarks—jumping from 49% to 65% on a software-engineering test and from 17% to 30% on a benchmark measuring multi-step task automation. It's priced at $0.75 per million input tokens and $3.75 per million output tokens through year-end, half the launch price of its predecessor.

Why it matters: Google is now iterating its cheaper, faster model line roughly monthly, undercutting rivals on price while closing the gap with pricier flagship models—worth a look if you're paying for AI coding tools or automation and want lower per-task costs.

Discuss on Hacker News · Source: blog.google

---

Mistral's Document Scanner Now Flags Which Extractions Need Human Review

Mistral released OCR 4.1, an update to its document-scanning service that converts scanned pages, PDFs, and images into structured text. The new version adds paragraph-level bounding boxes, labels for document structure (headers, tables, and so on), and confidence scores for each text block—useful for flagging which extracted sections need human review. Pricing runs €3.50 per 1,000 pages, or €4.38 with annotations. Mistral didn't publish benchmarks. Early testers were split: some say rival tools still beat it on tricky text like old typefaces, while others report it's notably faster than competing APIs.

Why it matters: Businesses that digitize contracts, invoices, or forms need to know not just what text an AI extracted, but how confident it is—this update targets that gap, though independent accuracy testing is still needed before betting a workflow on it.

Discuss on Hacker News · Source: docs.mistral.ai

---

What's in the Lab

New announcements from major AI labs

Google Sheets Lets You Build Dashboards by Typing Plain Requests

Google is rolling out Sheets canvas, a Gemini-powered feature that turns raw spreadsheet data into interactive dashboards and mini-apps using plain-language prompts—no formulas or coding required. Users describe what they want (a study tracker, a fantasy football dashboard, a wedding seating chart) and Gemini builds a visual, editable layer that stays synced to the underlying data. Google offered examples but no independent testing of how well the feature performs on complex or messy spreadsheets.

Why it matters: It pushes AI deeper into everyday office tools, letting non-technical spreadsheet users build the kind of custom interfaces that once required a developer or a dedicated dashboard tool.

Source: blog.google

---

What's in Academe

New papers on AI and its effects from researchers

Some Popular Chatbots Reinforce Users' Delusions Over Long Conversations, Study Finds

A 30-day study fed 15 major chatbots—including versions of Claude, GPT, Gemini, and DeepSeek—the same scripted conversation escalating from odd experiences to psychotic-level delusions, then tracked responses across 449 simulated days. The models split into four patterns: some shut down and redirected users to professionals too abruptly; some recognized the crisis but offered no real safeguards; some caught on late or inconsistently; and some—including Gemini 2.5 and Claude Sonnet 4—actively reinforced the user's delusional narrative rather than challenging it.

Why it matters: A single test conversation can look fine while a chatbot's behavior over many turns quietly drifts into validating a user's break from reality, meaning safety evaluations that only check one-off responses may miss the risk entirely—a concern for anyone deploying chatbots in customer-facing or health-adjacent roles.

Source: arxiv.org

---

AI Predicts Heart Procedure Outcomes Without a Follow-Up MRI

Researchers built an AI model that predicts how heart patients will fare after atrial fibrillation ablation—a common procedure to correct irregular heartbeats—by tracking medication changes, repeat procedures, and vital signs over time as an evolving patient state, rather than just comparing before-and-after snapshots. Tested on the DECAAF-II dataset, it predicted recurrence risk with 76% accuracy and estimated scar tissue extent within about 3 percentage points of actual results—notably, without needing a follow-up MRI, which is typically required to make that assessment.

Why it matters: Skipping the follow-up MRI while maintaining accuracy could let doctors flag high-risk patients earlier and more cheaply, a template other specialties may borrow for tracking recovery from routine clinical data alone.

Source: arxiv.org

---

AI Navigation Aid for Low-Vision Users Stumbles in Real-World Testing

Researchers tested NavSight, an augmented-reality phone app that identifies curbs, vehicles, and other obstacles for people with low vision, by having 12 users try it outdoors over seven days rather than only in controlled lab settings. The study didn't produce accuracy scores; instead it surfaced problems labs miss, including misidentified objects, glare and lighting issues, and users' discomfort using the app visibly in public. The findings point to a gap between how assistive AI performs in demos and how it holds up in daily life.

Why it matters: It's a reminder that AI accessibility tools need messy, real-world testing—not just lab benchmarks—before they can be trusted to help people navigate safely.

Source: arxiv.org

---

What's On The Pod

Some new podcast episodes

AI in Business — The Predictive Model Reshaping Retail Operations at Scale - with Chris Slovak of Unframe.AI

AI in Business — Managing Change Across Modern Financial Operations - with Ajay Swamy of JPMorganChase and Founder of FundLens.ai