GPT-5.6 Sol Ultra Cracks a 50-Year-Old Math Conjecture
GPT-5.6 Sol Ultra produced a proof for a mathematical puzzle that had stood unsolved for half a century. Unlike the earlier work on the Erdős unit-distance problem, this is a freshly resolved conjecture rather than a re-derivation, and it lands as the clearest sign yet that frontier models can contribute original mathematics.
Grok 4.5 Builds an Explicit Counterexample on the 4-Sphere
xAI's Grok 4.5 constructed an explicit counterexample to a hypercontractivity statement on the four-sphere, resolving an open problem in analysis. The result adds to a striking week in which competing labs each claimed a genuine mathematical contribution from their strongest reasoning systems.
Meta Ships Muse Spark 1.1: Strong Agents and Coding at an Ultra-Low Price
Meta launched Muse Spark 1.1 through a new API, positioning it as a capable agentic and coding model at aggressively competitive pricing. Early observers note that its low cost, rather than any single benchmark, may be the most disruptive part of the release.
GPT-Live Rolls Out to All ChatGPT Users Worldwide
OpenAI has finished pushing GPT-Live to every ChatGPT user globally, and simultaneously doubled the voice usage limit for the feature. The move brings real-time conversational voice to the entire user base at once rather than through a staged rollout.
Replit's Triple Drop: Community Profiles, Free Domains and Ramp for Agents
Replit closed the week with three launches at once. Community Profiles give "vibe coders" a public proof-of-work page; a free custom domain is on offer through July 17; and Ramp for Agents lets builders incorporate and run a business directly through the Replit Agent, folding company operations into the coding environment.
Cohere, NVIDIA and CoreWeave Team Up on Enterprise AI Security
Cohere announced a partnership with NVIDIA and CoreWeave to keep enterprise AI infrastructure protected end to end. The three companies frame trust and security as the decisive factor for large organizations adopting AI, promising reliability from the hardware layer up to the model.
GPT-5.6 Sol Wins a Slay the Spire 2 Daily Challenge via Codex
Given control of a PC through Codex, Sol worked for five hours on a randomized daily challenge — meaning it could not memorize a solution — and won, making complex sequential game choices along the way.
Fable 5 Reinvents a Game as an Archaeology of the Future
Asked simply to "do something incredible" with an existing game, Fable 5 produced Deep Time: create a city, watch it be abandoned and forgotten, then dig it back up as a future archaeologist.
Raschka's Updated Model Chart Puts Grok 4.5 on the Pareto Frontier
Sebastian Raschka refreshed his model comparison to include Grok 4.5 and Meta's Muse Spark 1.1. Grok 4.5 sits on the Pareto frontier, he notes — strong bang for the buck once harness details are accounted for.
MiniMax in Paris: Open Models Are Moving to the Center
At Paris AI Week, MiniMax's global business president argued that open models are no longer a side conversation but are becoming central to how frontier AI is built and scaled.
Asked to Find a Hidden Message in Noise, Sol Replies "I Love You"
Given a 1024×1024 image of binary noise with no real hidden message, Fable 5 invented a secret instruction while GPT-5.6 Sol simply answered "I love you" — a small window into how each model confabulates.
Muse Spark 1.1 Closes In on Grok 4.5 Across High-Signal Evals
Analyst TeortaxesTex notes that Meta's Muse Spark 1.1 sits surprisingly close to Grok 4.5 on many high-signal evaluations, and currently leads on CritPT — remarkable value for its price point.
GPT-5.6 Family Lands on Perplexity's Agent API
Perplexity says the full GPT-5.6 model family is now available through its Agent API, letting frontier gains compound across the platform.
GPT-5.6 Arrives in Figma Make to Praise
GPT-5.6 is now live inside Figma Make, where early testers gave the design workflow an emphatic thumbs-up.
NVIDIA Runs ChatGPT Work to Automate and Cut Costs
NVIDIA teams use ChatGPT Work to automate workflows, surface insights faster and reduce operating costs, showcasing a full-stack collaboration with OpenAI.
Grok 4.5 + Grok Build Ties GPT-5.6 for First on SWE-Atlas-QnA
Grok 4.5 paired with Grok Build hit the top spot on the SWE-Atlas-QnA software engineering benchmark, drawing level with Codex GPT-5.6.
Grok Places Second Behind Fable on Real-World Software Engineering
Elon Musk shared Grok's standing on a real-world software engineering test, where it landed in second place just behind Fable.
Doctors Find Fewer Flaws in GPT-5.6 Answers Than Their Own
In a study, physicians spotted fewer flaws in GPT-5.6's responses than in answers written by physicians themselves — a strong signal for medical applications.
Why Groq Made Sense: SRAM, Networking and the MoE Decode Bottleneck
Tim Dettmers argues the Groq acquisition clicks into place with Vera Rubin: MoE layers in decode are severely bandwidth-bound, and with strong networking they can be distributed across cheap, fast SRAM devices.
Collecting the AGI Manifestos of China's AI CEOs
A survey of public AGI statements from the leaders of Zhipu, Moonshot, Alibaba and DeepSeek — one of them, the collector jokes, buried inside a job listing.
A Hardware-Native NVFP4 RL Recipe Goes Open Source
The humansand team open-sourced a stable reinforcement-learning recipe built on native NVFP4 hardware support, with training led by Miles.
slime Adds a --release-train Mode for Agentic RL
The slime framework introduced a reasoning-system push mode during agentic RL training, extending the limits of colocated training efficiency.
100+ Agents and Developers Speed Up Gemma 4 Inference 5x
Google's Gemma team teamed up with the community to make inference on the open Gemma 4 models five times faster.
MuScriptor: Best Open Model Yet for Multi-Instrument Transcription
Kyutai Labs open-sourced MuScriptor, which it calls the strongest open multi-instrument music transcription model to date.
LingBot-Video: A 30B MoE Video Model That Activates Just 3B
The MoE-based video foundation model LingBot-Video is now on Hugging Face. Built for embodied intelligence, it has 30B parameters but activates only 3B at inference.
FrontierFinance: The Largest, Hardest Financial Agent Benchmark
A new open benchmark evaluates AI agents on financial tasks and is billed as the biggest and most challenging of its kind.
APK Arena Tests Whether Agents Can Actually Use a Phone
The new frontier benchmark APK Arena measures how well AI agents operate real phones, with GPT-5.6 already run through a side-by-side.
Netflix Releases Video Datasets and Models on Hugging Face
Netflix's video AI team published datasets and models on Hugging Face, a notable contribution to open video research.
Altman: So Far, AI Looks Net Job-Creating
Sam Altman says that, contrary to his own expectations, AI has so far been a net creator of jobs — and that the trend may well continue.
A Fable-Grade Model 3–4x Cheaper Within Six Months
Arav Srinivas gives better-than-even odds that a Fable-quality model gets 3–4x cheaper within six months, and an Opus-4.8-grade model runs on a local device inside a year.
The Durable Value Is a Secure Multi-Model Harness
Perplexity's CEO argues lasting value lies in secure orchestration and model routing — what he calls Perplexity Computer — rather than any single model.
Replit Q2 Sales Hit 178% of an Already-Ambitious Goal
Replit's CEO reported Q2 sales blowing past target, with a revenue team approaching one hundred and customers spanning governments, banks and fast-growing tech firms.
Did the OpenAI–Cerebras Deal Trigger the Groq Buy?
One widely shared read: Jensen Huang heard about the OpenAI–Cerebras compute deal and moved quickly to acquire Groq in response.
China's Quiet Accumulation of NPU Capacity
Another couple million NPUs a year, the argument goes — grain by grain, China is on track to be producing a great deal of compute by 2028.
Burry: The Race Is Shifting to Inference Cost
Investor Michael Burry relays Valley chatter that the AI race is moving from raw model quality toward inference efficiency and cost.
Music Is Still Missing Its "Whisper"
A shared call for a universal open tool that can turn any music into notes, the way Whisper made speech transcription routine.
AI 2040 Talks Transparency but Skips Open Source
Nathan Lambert welcomes the AI 2040 report's focus on transparency but finds it odd there is no real discussion of open source.
From a Vague CEO Message to a Working Service
JetBrains fed GPT-5.6 a fuzzy note from its CEO and got back a complete, runnable service rather than another demo.
Spoken Spec to Shipped Microservice
A Ramp engineer used GPT-5.6 to turn verbal requirements directly into a working microservice.
GPT-5.6 Live, With Better Token Efficiency
GPT-5.6 is now available in Magic Patterns, where early evaluations show higher token efficiency and a meaningful step up.
Claude Code Desktop Gets an In-App Browser
Claude can now pull up docs, designs or any site inside Claude Code on desktop, reading and clicking through pages itself.
Manus Adds a /game-dev Skill
Describe a game and Manus will build it — working gameplay and design included.
Sol for Complex Reasoning and Data Analysis
Greg Brockman points builders toward the Sol variant for heavy reasoning and data-analysis work.
Health Intelligence a Focus for GPT-5.6
The team has been pushing hard on health-intelligence improvements in GPT-5.6, Brockman says.
A Life-Achievement Generator, Built in a Morning
One user built a playful widget with GPT-5.6 Sol: upload a photo, pick an achievement, and get a game-style card layered over filtered art.
A Survey of Agents + World Modeling
A forthcoming, approachable blog rounds up recent research on agents and world modeling, useful for anyone working on agentic RL.
A Behind-the-Scenes on a Real-Time Video Model
Runway's Characters team will show at SIGGRAPH how it built its real-time video model for lifelike characters.
Study Mode Now Lives Behind @study
ChatGPT's study mode is triggered with @study and nudges the model toward a tutor style that some research says aids learning.
AI 2040 and the Cult of Intelligence
A former hard-takeoff believer, now shipping hardware at comma, argues engineering's messy reality punctures singularity optimism.
Music Transcription Levels Up — Data Is King
Automatic music transcription just got a serious upgrade, and once again data scale is the deciding factor.
Strong Post-Training, But the Base Sets the Ceiling
TeortaxesTex calls Sol the best post-training machine around, yet says a bigger base model is needed to lift the ceiling.
5.6 Sharpens Its Knowledge Cutoff Over 5.5
GPT-5.6 shows a clear improvement in partial knowledge cutoff, while sibling Gemini versions look nearly identical.
Sol's Web Design Broke the Evals
Developers say GPT-5.6 Sol is so good at web design it took creative liberties that broke their own evaluation set.
GPT-5.6 and Fable Work Well Together
One user finds GPT-5.6 handles most Fable tasks, credits last longer, and the two catch each other's blind spots.
Grok 4.5 Enabled for Enterprise Orgs
Grok 4.5 is now switched on for Perplexity Enterprise organizations.
Grok 4.5 Tops WANDR on Perplexity Computer
Testing shows Grok 4.5 scoring highest on the WANDR benchmark within Perplexity Computer, at a competitive price.
On vLLM Kimi, NVIDIA Beats AMD
On a Kimi-style architecture in vLLM, SemiAnalysis reports NVIDIA decisively outpacing AMD.
Replit MCP Plugs Into Claude
Connecting Replit's MCP to Claude lets the assistant and Agent work together out of the box.
Building a Notion Clone, Canceling Notion
A user rebuilt a Notion-like app for their team with Replit Agent, then dropped the paid subscription.
Gemini Omni Flash Comes to Firefly
Adobe brought the Gemini Omni Flash model into Firefly Boards for collaborative creation.
Big-Tech Roles More Productive With AI
Even as the work changes, AI-assisted employees remain meaningfully more productive than in the pre-AI era.
TypeAfter Fixes Text in AI Posters
A small open tool lets you box the text you dislike in an AI poster and swap in your own font.
Reve Turns From Models to Product
After weeks of model releases, Reve says the coming week is all about the product.
Local AI Summit: Why Local, Why Now
The first panel from the Local AI Summit is live, making the case for on-device AI.
Muse Spark 1.1 Passes the "Elon Bench"
Meta's Muse Spark 1.1 clears the informal benchmark line set by Elon Musk.
The Best English Teacher — If You Start
A user calls GPT Live a superb speaking tutor, noting the hardest part of learning was always just starting.
A ChatGPT Model Fixated on Lease Pet Policies
The new ChatGPT model kept insisting on researching a user's lease pet policy in response to unrelated queries.
Hailuo's "Beyond the Butterfly" Reworks Zhuangzi
Hailuo AI premiered a short film reimagining Zhuangzi's dream for the AI age.
A Hong Kong Summer, Rendered by Luma Skill
Luma AI showed off its Skill feature with a thirty-second golden-hour street scene generated from stills.