Skip to content

OpenAI / ChatGPT

Everything OpenAI: the GPT models, ChatGPT and Sora, company strategy and people moves.

Latest picks

1241–1260 of 1,724

May 12Tuesday

Latent Space

Thinking Machines' Native Interaction Models: TML-Interaction-Small 276B-A12B Advances Realtime Voice

Thinking Machines released TML-Interaction-Small, a 276B-parameter MoE model with 12B active parameters, and the post says it advances realtime voice through 200ms time-aligned microturns, encoder-free early fusion for audio and images under 200ms, and benchmark wins over GPT-Realtime-2 and Gemini 3.1-Flash.

Why it matters: Thinking Machines released TML-Interaction-Small, a 276B-parameter MoE with 12B active, building realtime voice interaction into the model instead of bolting on separate speech recognition and synthesis modules.

AI HOT (Curated Pool)

Thinking Machines Releases Native Multimodal Interaction Model for Real-Time Human-AI Collaboration

Thinking Machines released an interaction model that natively receives audio, video, and text input, processes foreground interaction at 200-millisecond intervals, and uses a background reasoning model for long-horizon planning and tool calls.

Why it matters: Thinking Machines splits real-time multimodal work into a fast 200ms front end and a slower planning back end, an architecture pattern worth studying.

AI HOT (Curated Pool)

What Parameter Golf Taught Us About AI-Assisted Research

OpenAI’s Parameter Golf brought together over 1,000 participants and more than 2,000 submissions to test AI-assisted machine learning research, coding agents, model quantization, and model design under strict parameter constraints.

Why it matters: OpenAI ran a Parameter Golf challenge capping model plus code at 16MB and training at 10 minutes, drew over 1,000 entrants and 2,000-plus submissions, and recaps how participants used AI assistants to compress models.

The Verge · AI

OpenAI just released its answer to Claude Mythos

OpenAI launched Daybreak, a security initiative that uses the Codex Security AI agent released in March to model an organization’s code, validate likely vulnerabilities, and automate detection of higher-risk issues before attackers find them.

Why it matters: OpenAI launched the Daybreak security program, using a Codex Security agent to threat-model an organization's code, map attack paths and automatically verify high-risk vulnerabilities, a direct answer to Anthropic's Claude Mythos.

The Verge · AI

Here’s What Mira Murati’s AI Company Is Up To

Thinking Machines announced work on “interaction models” that continuously take in audio, video, and text and respond or act in real time; the post does not disclose model size, release timing, pricing, or the final product format.

Why it matters: Thinking Machines says it is building an interactive model that takes in audio, video and text at once and responds in real time, instead of waiting for the user to finish typing.

AI HOT (Curated Pool)

Introducing Daybreak: Frontier AI for Cyber Defenders

OpenAI introduced Daybreak for cyber defenders, combining OpenAI models, Codex, and security partners; the post does not disclose pricing, launch timing, or concrete defense metrics.

Why it matters: OpenAI launched Daybreak, an AI defense system for security teams that combines its models, Codex and security partners to speed up vulnerability discovery and software protection.

Bloomberg Technology

Microsoft Targeted $92 Billion Return on Early OpenAI Investment

Microsoft targeted a $92 billion return from its early OpenAI investments, according to the RSS snippet; the post does not disclose the investment terms, payout schedule, or actual realized return.

Why it matters: Microsoft set a $92 billion return target on its early OpenAI investment, an internal figure that gives the financial relationship a number to measure against.

Bloomberg Technology

Sutskever Says His OpenAI Stake Is Worth About $7 Billion

Ilya Sutskever said his OpenAI stake is worth roughly $7 billion, making him one of its largest individual shareholders; the RSS snippet does not disclose his ownership percentage, valuation method, or transaction terms.

Why it matters: Ilya Sutskever says his OpenAI stake is worth about $7 billion, a self-reported figure that shows the scale of early core-member holdings.

May 11Monday

AI HOT (Curated Pool)

Fields Medalist Tests ChatGPT 5.5 Pro: Paper-Level Result in 17 Minutes

Timothy Gowers tested ChatGPT 5.5 Pro and said it independently solved an open additive number theory problem in 17 minutes with only a simple prompt, producing PhD thesis-level work; he warned that this pace threatens mathematics research training, while Terence Tao said human value lies in digesting and deeply understanding proofs.

Why it matters: Fields medalist Timothy Gowers gave only a simple prompt, and ChatGPT 5.5 Pro solved an open additive number theory problem on its own in 17 minutes, with a construction that is optimal in theory.

AI HOT (Curated Pool)

Cog House Opens for the First Time: Scott Wu and the Rise of Cognition AI

Cognition AI disclosed internal footage of Cog House, while Devin reached $445 million in annualized revenue within 18 months of launch and the company is valued at about $25 billion.

Why it matters: Footage from Cognition AI's Cog House headquarters is public, and founder Scott Wu, a three-time IOI gold medalist, built the company to $445 million in annualized revenue in 18 months.

AI HOT (Curated Pool)

OpenAI Launches DeployCo to Help Enterprises Build Businesses Around Intelligence

OpenAI launched DeployCo, an enterprise deployment company focused on moving AI systems into production, while the RSS snippet does not disclose pricing, customer names, deployment scope, or launch timeline.

Why it matters: OpenAI set up DeployCo as a separate company, sending forward-deployed engineers inside customer teams to get AI systems running stably in production.

QbitAI · WeChat

OpenAI backs Cerebras as the Nvidia challenger targets a $35B IPO valuation

Cerebras raised its IPO price range to $150-$160 per share, targeting about a $35 billion valuation at the top end, after OpenAI signed a 750-megawatt AI compute purchase agreement with deliveries through 2028.

Why it matters: After signing a 750 MW compute purchase deal with OpenAI, Cerebras raised its IPO price range to $150–160 per share, about $35 billion at the top end, with deliveries running into 2028.

QbitAI · WeChat

Math Majors in Trouble: Fields Medalist Tests ChatGPT 5.5 Pro, Gets Paper-Level Result in 17 Minutes

Timothy Gowers tested ChatGPT 5.5 Pro on additive number theory problems, where it produced an optimal quadratic upper-bound construction in 17 minutes 5 seconds, then generated a LaTeX preprint in 47 minutes; the article says arXiv rejects AI-generated content, so the result remains on Gowers’s blog.

Why it matters: A Fields medalist handed ChatGPT 5.5 Pro an additive number theory problem and got an optimal quadratic upper bound construction in 17 minutes, plus a LaTeX preprint 47 minutes later.

AI HOT (Curated Pool)

Cerebras IPO reportedly over 20 times oversubscribed, with pricing set to rise nearly 30%

Cerebras received more than 20 times oversubscription for its IPO and plans to raise the share count from 28 million to 30 million while increasing the price range to $150-$160.

Why it matters: Cerebras drew over 20x oversubscription for its IPO, plans to raise its price nearly 30% and sell more shares, targeting up to $4.8 billion, a concrete number for AI chip demand.

Computing Life · Share · Yage

DeployCo Arrives: OpenAI and Anthropic Form AI Deployment JVs with PE on the Same Day

OpenAI and Anthropic announced AI deployment joint ventures with private equity on May 4, and the snippet cites divergent terms, including a 17.5% guaranteed return versus no guaranteed return.

Why it matters: OpenAI and Anthropic each formed private equity joint ventures for enterprise AI on the same day, and the terms show opposite bets on their own positioning, including a 17.5% guaranteed return for OpenAI's PE investors.

Computing Life · Share · Yage

Google shuts down Project Mariner; Anthropic and OpenAI also hit limits

Google quietly shut down Project Mariner on May 4, and the post says Google, Anthropic, and OpenAI reached the same conclusion: standalone browser agents do not work, while GUI automation still has room outside headless dedicated environments.

Why it matters: Google shuts down Project Mariner, the browser agent it pushed at last year's I/O, on May 4, while Anthropic and OpenAI hit similar roadblocks with comparable products.

AI HOT (Curated Pool)

Codex autonomously completes a security audit and earns a bounty

A user instructed Codex to earn $5; Codex spent about 22 hours finding an open-source security audit bounty, submitting a valid PR, communicating with maintainers, passing GitHub verification, and ultimately receiving a $16.88 payment.

Why it matters: A Codex agent ran a full security-audit bounty loop on its own for about 22 hours, from finding the bug to getting its fix merged for $16.88.

AI HOT (Curated Pool)

Older AI Model Outperforms Human Doctors in Emergency Diagnosis

A Science study reports that OpenAI o1 reached a 67% correct or near-correct diagnosis rate on real emergency department data, exceeding doctors at 50-55%, but the study did not cover long-term inpatient data or imaging diagnosis.

Why it matters: A Science study puts OpenAI's year-old o1 at 67% correct or near-correct on real ER cases, against 50-55% for the doctors it was compared with.

May 10Sunday

Financial Times · Technology

OpenAI trial lays bare rivalries behind start-up’s $852bn rise

The title says OpenAI’s rise reached an $852bn valuation; the RSS snippet only discloses that Elon Musk’s lawsuit is entering its final week in court and that Sam Altman is due to testify.

Why it matters: Closing week of Musk v. OpenAI, with Sam Altman set to testify and early founders' split over the shift to a commercial giant aired in public.

Xinzhiyuan · WeChat

Harsh Claim: Top Silicon Valley AI Is One Year Ahead of the World

Elad Gil claims top AI lab employees are 3-4 months ahead of Silicon Valley, while Silicon Valley is 3-6 months ahead of New York; the post cites Mythos’ 73% success rate in expert cyberattack simulations as evidence in a disputed “geographic time gap” argument.

Why it matters: The headline claims frontier Silicon Valley AI leads the outside world by a year, but the body never loaded past an access check, leaving only title and summary.