In partnership with
In Todayโs Issue:
๐๏ธ ChatGPT learns to listen and talk at the same time
๐งช OpenAI says a third of a top coding test is broken
๐ธ Executives are rattled by their AI bills
๐ Grok 4.5 goes cheap and climbs the coding charts
โจ And more AI goodnessโฆ
โก The Signal
The biggest AI voice release of the year is a bet that how present a model feels matters more than how smart it tests.
OpenAI's GPT-Live drops the walkie-talkie rhythm that made voice assistants feel robotic and lets ChatGPT listen and speak at the same moment, murmuring "mhmm" while you talk and quietly handing the heavy reasoning to GPT-5.5 in the background. The wager is that feeling present, more than topping another benchmark, is what finally gets people using voice AI every day. The same day showed the other face of the race: xAI's Grok 4.5 landed mid-pack on general intelligence yet near the very top on coding agents, at a fraction of the leaders' price. And under both launches sits the unglamorous plumbing, from Beijing rationing Nvidia's H200 chips to executives quietly panicking over their AI invoices. Capability is getting cheap fast, and trust, cost control, and human attention are becoming the real bottlenecks.
All the best,
Kim Isenberg
๐ง New Podcast Release
|
๐งช OpenAI Says a Third of a Top Coding Test Is Broken
OpenAI just told the industry to stop trusting the coding benchmark it recently championed. After auditing SWE-Bench Pro, a test built to measure how well AI agents handle realistic, long-horizon software tasks, OpenAI found that between 27% and 34% of its tasks are broken, with buggy setups or impossible grading, and it publicly retracted its own earlier call to adopt it. An automated pass flagged 200 tasks; five human software engineers flagged 249.
๐ tl;dr: When the referees are broken, every leaderboard built on them is suspect.
๐ธ Executives Are Getting Sticker Shock From AI
The pitch was cheap AI that quietly replaces workers; the invoice is turning out to be the shock. A KPMG survey of 2,145 senior executives across 20 countries found that 29% have no idea where their AI costs are even coming from, and a third say their own lack of AI knowledge is slowing rollouts. As vendors shift from flat-rate deals to usage-based pricing, spending that used to be subsidized now lands straight on the budget.
๐ tl;dr: Adoption was the easy part; forecasting the bill is the hard part.
๐ China Cracks the Door Open for Nvidia's H200
Beijing is preparing to let its biggest AI firms buy a limited batch of Nvidia's H200 chips, reversing months of resistance. According to The Information, Chinese officials have told Alibaba, ByteDance and DeepSeek they may soon win approval, with strings attached: the chips are for training only, inference should stay on domestic processors, and the total could be capped under 200,000 units, less than half of what the firms requested. It reads as a pragmatic retreat forced by a real compute shortage, not a change of heart.
๐ tl;dr: China still wants Nvidia's chips; it just wants them on its own terms.
Put GPT-Live's new full-duplex voice mode to work as a live thinking partner rather than a hands-free search box.
Why it helps: Because it keeps listening while you talk, you can reason out loud and let it cut in the instant something does not add up, the way a sharp colleague would.
Try this: Open ChatGPT voice, switch on GPT-Live, and say: "I'm going to think out loud about a decision I'm stuck on. Don't wait for me to finish. Interrupt me the moment you hear a shaky assumption, a missing option, or a claim I can't back up, and push me to get specific."
๐ฌ Watch This
OpenAI's own team demos GPT-Live, the full-duplex voice model behind today's Featured Story.
Researchers Kundan Kumar, Yuchen Zhang, Ehsan Asdar and Rithesh Kumar run live, unscripted conversations where the model listens and talks at the same time, handles interruptions without losing its place, switches languages mid-sentence, and quietly hands the harder questions to GPT-5.5 on the fly.
It is the clearest look yet at the whole pitch: a model that feels like it is actually listening rather than waiting for its turn to speak.
"I'll refer to this as the 'compressed 21st century': the idea that after powerful AI is developed, we will in a few years make all the progress in biology and medicine that we would have made in the whole 21st century."
โ Dario Amodei, CEO of Anthropic, โMachines of Loving Graceโ
The rumour mill has already moved past GPT-5.6 to something bigger. According to chatter from accounts that claim inside knowledge, OpenAI's next major model, which some are calling GPT-6, is said to be a much larger, freshly pretrained system built as its true answer to Anthropic's "Mythos," and it reportedly arrives far sooner than most expect, possibly within weeks. The same sources suggest any release could be held back at first, echoing how recent frontier models have been gated. They also claim everyone is scaling up at once, with Elon Musk's xAI said to be training a 10-trillion-parameter Grok, and both leading labs reportedly seeing "nothing above us but air." Treat all of it as unverified: these are anonymous claims, not announcements, and AI release timelines have a habit of slipping.
The End of Taking Turns With AI
The Takeaway
๐ OpenAI launched GPT-Live and GPT-Live-1 mini, voice models that listen and speak at the same time instead of taking turns.
๐ For hard questions the model hands off in the background to GPT-5.5, so quick chat and slow reasoning run on separate tracks.
๐ On BrowseComp, a web-search test, GPT-Live-1 scored 75.2% against just 0.7% for the old Advanced Voice Mode.
๐ It is rolling out worldwide now on iOS, Android and the web, but cannot yet do video or screen sharing.
For three years we have talked to AI on the machine's terms: type a prompt, wait, read the reply, repeat. GPT-Live is OpenAI's bid to flip that and let us deal with AI the way we deal with people. The old voice mode worked like a walkie-talkie: you spoke, it waited for silence, then it answered. GPT-Live uses a full-duplex design that processes audio continuously and decides many times per second whether to talk, listen, pause or jump in. It can murmur "mhmm" while you are still speaking, let you cut it off mid-sentence, or sit quietly when you need a beat to think.
(The Decoder)
The shift works because OpenAI split talking from thinking. GPT-Live runs the live conversation itself, but hands anything that needs a web search or deeper reasoning to GPT-5.5 in the background, then folds the answer back in. That is how it keeps a fast, natural back-and-forth going while still doing real work: on BrowseComp, a web-search benchmark, GPT-Live-1 scored 75.2% against just 0.7% for the old Advanced Voice Mode, and on graduate-level science questions (GPQA) it reached 84.2% at its highest reasoning setting.
(The Decoder)
Changing how we interact with AI cuts both ways. GPT-Live still cannot do video, screen sharing or full multilingual parity, and early users say it can be over-eager, jumping in when they wanted silence. The deeper catch is that a more human interface pulls on human instincts: OpenAI's own research notes that lifelike voices deepen how attached people feel to a model and sharpen its powers of persuasion. The more talking to AI feels like talking to a person, the easier it is to forget that it is not one.
Why it matters: The interface is becoming the product. Once talking to AI feels like talking to a person, voice stops being a feature and turns into the default way most people reach for it, which makes how these systems behave in conversation, and how far we trust them, the thing that matters most.
Own AI deployment, grow your career
Making AI actually work day to day is becoming its own job.
On July 16, hear from three people doing it: Simone Santiago Broad (Yoco), Yelva Espinoza (Zumba Fitness), and Fin's Dave Lynch. They'll share what the role really looks like, how it came to exist, the skills worth hiring for, and the challenges they're tackling right now. Bring your questions, since the best moments happen live.
Register to save your spot.
The chart: Artificial Analysis benchmarked xAI's new Grok 4.5 (high) two ways. On its headline Intelligence Index, a composite of nine reasoning, knowledge, math and coding tests, Grok 4.5 scores 54, landing fourth behind Claude Fable 5 (60), Claude Opus 4.8 (56) and GPT-5.5 (55). But on the Coding Agent Index it jumps to 76, tied for second and one point off the leader, Claude's Fable 5 (77).
The lesson: Grok 4.5 is a specialist punching above its general rank. It was trained alongside the Cursor coding editor, and it shows: on agentic coding it matches models that beat it on raw intelligence, while costing $2 per million input tokens against far pricier frontier rivals.
The caveat: These are "high" reasoning-effort scores, and a coding-agent composite is not the same as everyday reliability. A leaderboard also cannot see the data-quality problems OpenAI flagged in SWE-Bench Pro today, a reminder to read every ranking with a skeptical eye.
๐ง Your Brain Knows the Difference Between Scrolling and Learning
โก Bottom line: A new brain-imaging study finds people feel they are learning from short videos while remembering measurably less.
๐ก Why it matters: The feeds that dominate our attention, tuned by recommendation AI, may be quietly training us to absorb less.
๐ What it means: As AI floods the internet with bite-sized clips, the format itself, as much as the topic, shapes what sticks.
Everyone has felt it: an hour of short clips flies by and leaves almost nothing behind. A new study in Communications Psychology puts hard evidence under that hunch. Researchers at Yunnan Normal University ran three experiments with hundreds of students and found that people who watched information chopped into rapid short-form clips scored worse on quizzes than those who watched the same material as one continuous video, even when they tried hard to learn.
(PsyPost)
The brain scans showed why. During fragmented viewing, activity across memory-related regions fell out of sync and the brain slipped into a kind of "sensory tracking," following the stream of images instead of doing the deeper work of forming memories. Yet the students still felt like they were learning. That gap between the feeling of learning and the fact of it is the illusion in the study's title.
(The Decoder)
This is an AI story through the back door. The short-video feeds that now shape how billions of people take in information are curated by recommendation algorithms tuned for watch time rather than retention, and the next wave of that content will be AI-generated at effectively infinite scale. For a readership betting on AI to extend healthy, capable human lifespans, cognitive health belongs on the list. Guarding your attention may turn out to be as much a health habit as sleep or exercise.
Why did one company's AI work, and another's didn't?
One had a dedicated owner. Resolution rate: 48.9%. One didn't: 0.38%. See the full breakdown.
