SShortSingh.
TopTechnologyAI & MLStartupsProgrammingScienceBusinessFinanceCrypto & Web3World
Back to feed
Programming

Inception Unveils Mercury 3, a Diffusion-Based LLM Claiming 1,000 Tokens per Second

0
·1 views

Inception has introduced Mercury 3, a large language model built on diffusion techniques rather than the conventional autoregressive approach. The company reports that the model can generate output at roughly 1,000 tokens per second. The announcement was shared on Hacker News, where it drew 63 points and 18 comments from readers. The source material provides only limited detail beyond the headline claim, so specifics on benchmarks and methodology are not confirmed here.

Read the full story at Hacker News

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Log in to join the discussion and vote.

Log in

Related stories

0
ProgrammingHacker News ·

Open-source phonics reader lets early readers sound out words in stories

A parent built a free, open-source book reader app to help their six-year-old daughter practice reading beyond repetitive phonics lessons. The app lets readers tap words to hear each sound and sound them out independently. A twelve-year-old sibling wrote the first story, which ends on a cliffhanger, and more chapters and stories are planned. The reader also supports an MCP connection that lets AI assistants import a user's own owned books from PDFs or phone scans into a private account. The included stories are released under Creative Commons, and the code is open source for anyone who wants to modify it.

0 comments Read more at Hacker News
0
ProgrammingDEV Community ·

Cheaper Alternatives to interviewing.io for Tech Interview Practice, Ranked by Budget

A DEV Community guide reviews alternatives to interviewing.io, whose live mock interviews start at $179 per session. It lists free peer-practice options such as Aced, a low-cost AI interviewer, and paid coaching through DesignGurus and IGotAnOffer. The author discloses building one of the listed tools, DevInterview.AI, and notes its limitations compared with human interviewers. The article recommends cheap, high-volume practice first, with one expert session shortly before a real onsite interview. All prices were stated as checked by the author on October 19, 2026.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Deterministic Replay Lets AI Agents Time-Travel Through Runs to Debug Distributed Systems

AI agents now write code faster than engineers can review it, and conventional diff inspection struggles with distributed state machine bugs and timing race conditions. Platforms such as Antithesis expose deterministic time-travel APIs that let agents step back through historical system runs, branch interactive terminals, and check fixes against real execution failures. The approach carries costs, since running systems inside specialized hypervisors requires strict containerized constraints. Unguided agents that branch through time can also consume large amounts of compute unless platform teams enforce hard runtime budgets. The article argues that platform engineering is moving from passive code review toward execution-proven verification, with the open question being whether these methods scale as microservice architectures grow more complex.

0 comments Read more at DEV Community
0
ProgrammingDEV Community ·

Open-source project lets a Safari agent control an iPhone via an ESP32 Bluetooth bridge

Developers have released iPhone Agent Harness, an MIT-licensed open-source project that lets users describe a phone task in Safari. A vision model inspects the screen while a nearby ESP32 board, connected over Bluetooth and Wi-Fi, supplies taps and keyboard input through AssistiveTouch. The demo shows the agent setting an alarm in the iPhone Clock app, though the authors note it is not a benchmark for arbitrary apps or long tasks. The setup requires iOS or iPadOS 27 or later, and firmware is installed through desktop Chrome or Edge. Users choose their own model provider, and a model can still misread the screen, so results must be checked on the phone.

0 comments Read more at DEV Community
Inception Unveils Mercury 3, a Diffusion-Based LLM Claiming 1,000 Tokens per Second · ShortSingh