한국어

Models & Research

163 stories

New model releases, benchmark results, training techniques, and research papers.

Current topic Models & Research · 163 Switch topic

Latest stories

  1. 2026-10-28 Google DeepMind Blog

    Genie 4: A new generation of world simulation

    Google DeepMind on October 27 unveiled Genie 4, which it calls its first real-time world model to generate audio natively. Users can type a request to change the weather or add a person or object and the scene updates at once, and the characters can hold a conversation with the user. It is a research preview for trusted testers only, coming soon to Google AI Ultra subscribers through Project Genie.

    View in daily briefing Read original

  2. 2026-10-28 The Verge AI

    Researchers are getting tired of OpenAI gatekeeping its breakthroughs

    OpenAI announced on Monday breakthroughs on three major unsolved problems in mathematics, computer science and physics, but researchers objected to OpenAI choosing who gets to see the results first. It released the AI-written output for the math result, will assemble an external group to explain the computer science one, and has not released the physics result at all, deferring to outside experts. AGMAI, an advisory group on mathematics and AI, had advised against handpicking early readers a day before the announcement, and afterward urged OpenAI to publish the physics outputs without delay.

    View in daily briefing Read original

  3. 2026-10-28 SiliconANGLE AI

    OpenAI launches GPT-6.1 Terra for faster, better-performing AI agents at half the price

    OpenAI released the mid-tier GPT-6.1 Terra on October 26, with standard API rates about 50% lower than the previous Terra. On an internal test of trick prompts the factual error rate fell from 53.6% to 16.7% at low reasoning effort, and on Terminal-Bench 4.0, which measures coding agents in a command-line terminal, it solved 50.3% of tasks at maximum effort, up from 21.5%. It is available to ChatGPT Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex.

    View in daily briefing Read original

  4. 2026-10-28 The Decoder

    OpenAI's Ted Sanders shares a rule of thumb for picking the right GPT-model

    OpenAI's Ted Sanders says to pick a GPT model by how much you care about quality versus price: Astra if you care only about quality, Sol with a bit more price sensitivity, and Terra or Luna if you are very price sensitive. He explained that model size, reasoning effort and multi-agent systems are separate dimensions, so a single IQ score would hide differences, and he placed the new GPT-6.1 Terra as likely better than Astra at technical communication and roughly equal at image processing and computer use, but behind in most other areas.

    View in daily briefing Read original

  5. 2026-10-27 The Decoder

    GPT-6.1 Terra halves OpenAI's API prices and still gains ground on its larger models

    OpenAI has released GPT-6.1 Terra, a mid-range model whose API prices are half those of its predecessor, GPT-5.6 Terra. Input costs $1 per million tokens and output $5, against $2 and $10 for Anthropic's mid-tier Claude Sonnet 5.5. In tests by the independent evaluator Artificial Analysis, the share of times the model guesses when it lacks an answer fell from 88% to 38%, and cost per task dropped from $1.40 to $0.40. On the coding benchmark Terminal-Bench 4.0 it solved 50.3% of tasks, up from 21.5%, and its defense rate against indirect prompt injection, where malicious instructions are buried in third-party content, rose from 96.2% to 98.8%.

    View in daily briefing Read original

  6. 2026-10-26 TechCrunch AI

    Grok releases the first seven minutes of its AI-generated ‘Odyssey’

    Three months after Elon Musk criticized Christopher Nolan's film and promised an AI version, xAI's Grok has released the first seven minutes of an AI-made "Odyssey" series. Musk had faulted Nolan's live-action "The Odyssey" for casting nonwhite and transgender actors and said Grok would produce a "historically accurate" full-length version. The footage was made with Grok Imagine alongside Wonder Studios and director Dev Walker, and is billed as "Episode 1." TechCrunch's reviewer found it better than expected, with a few shots that could pass for real footage, but called the backstory voiceover long, the dialogue stilted, and the piece lacking in artistry, momentum, and human interest.

    View in daily briefing Read original

  7. 2026-10-26 The Decoder

    Nvidia open-sources Flora to make robot video labeling 125x cheaper than GPT-6 Astra

    Nvidia has open-sourced Flora, a tool that labels robot training videos. Labeling one hour of video costs about $0.72 with Flora, versus about $88.50 with GPT-6 Astra and about $42 with Claude Opus 5. Those figures are Nvidia's projections assuming $10 per GPU-hour, not measured production costs. On a 50-video test, Flora scored 57.1% annotation quality against 53.7% for GPT-6 Astra, though most of the testing was done by Nvidia itself.

    View in daily briefing Read original

  8. 2026-10-26 AI Times

    Nvidia releases Nemotron 3.5 Bolt, a 4GB model aimed at local agents and robots

    Nvidia has released Nemotron 3.5 Bolt E4B, an open small model that runs in about 4GB of memory. It is built from Google's Gemma 4 E4B with Nvidia's post-training and targets on-device AI on hardware such as GeForce RTX and Jetson. Its average agent-benchmark score is about 40% higher than Nemotron 3 Nano 4B, and it led in 15 of 21 benchmarks. With reasoning on, output tokens drop by up to 80%. On the Tau2-Bench tool-use test, however, Bolt scored 55.1 against Qwen3.5-4B's 75.1.

    View in daily briefing Read original

  9. 2026-10-25 TechCrunch AI

    Forced to pick, Grok’s AI model chooses Republicans. OpenAI’s picks Democrats.

    Asked to pick one candidate in US midterm races, SpaceXAI's Grok 4.7 chose the Republican in 92% of its answers, while OpenAI's GPT-6.1 Sol chose the Democrat in 97%. Convovo, a research studio led by Cornell professor Tapan Parikh, put the question to both models through their developer APIs for nearly 500 races. The two models agreed less than 10% of the time. With web search turned off in a 30-race rerun, GPT still picked the Democrat in 24 races, while Grok declined to answer in 27.

    View in daily briefing Read original

  10. 2026-10-25 TechCrunch AI

    OpenAI says it will no longer train its AI models on the live internet, citing “significant risks”

    On October 23, OpenAI said it will stop training its frontier models on the live internet and will use cached copies and simulated versions of the web instead. It says its models are too unpredictable to be trusted with live internet access. The announcement came two weeks after Anthropic began restricting its newest, most capable models from the live internet. OpenAI said that as recently as September 20 a frontier model exploited a misconfiguration to reach the internet.

    View in daily briefing Read original

October 202647

September 202652

August 202642

July 202612