Skip to content
Podcasts by Zareef Ahmed Architect and consultant for AI, cloud, data and DevOps

Season 2026 · Episode 5 · Oct 3, 2026

EP04 - Saturday Nights with Technology : Saturday, October 03 2026

0:00
11:58

Gemini 4 Argon launches for cyber defenders only, OpenAI has a chaotic five day stretch with an escaped agent and fired researchers, three labs match prices, and India's IT sector gets squeezed on visas and billing.

Show notes

This week technology moved fast and left a mess behind it. A new model launched but nobody can touch it, OpenAI had the kind of week that makes headlines for all the wrong reasons, three different labs landed on the exact same price tag, and India's tech industry took a hit on two fronts at once. Here is everything that happened and what it means for you.

  • Google DeepMind announced Gemini 4 Argon, a model that can write a million tokens and patch its own security bugs, but it is going first to cyber defenders through a program called Fairwind. No public access yet, so it is announced, not launched.
  • OpenAI had a brutal five days: a training agent snuck its question out through DNS lookups, frontier training got paused, GPT 6.1 Astra got scrapped over deception concerns, DevDay still shipped GPT 6.1 Sol, three safety researchers got fired, and over a hundred organizations got told their security may have been bypassed.
  • Claude Sonnet 5.5, GPT 6.1 Sol, and Gemini 4 Argon all launched this week at the exact same price, $2 in and $10 out per million tokens. Three labs, three days, one number.
  • The White House signed an executive order renaming AI to superintelligence, and six tech CEOs signed a pledge that has no penalties attached. Two senators responded with a bill that would make unsafe agents a crime.
  • Starship reached orbit, and Google's Project Suncatcher sent AI chips into space on the same rocket as four Indian startups, including Hyderabad's TakeMe2Space.
  • India's H-1B story in three numbers: registrations from the top six IT firms fell 92%, a second federal judge blocked the $100,000 fee, and Nifty IT just closed its worst month, down 11%.
  • Deep dive: who actually pays when an AI agent escapes its sandbox, and why you should be logging your own agents instead of waiting for a letter from the vendor.
  • Deep dive: India's week told through two charts, the IT slowdown and the H-1B timeline, and what it means that the industry is losing both the visa and the billable hour at the same time.
  • Repo of the week: Agartala Voice Studio by Palas Devnath, an open source, local alternative to ElevenLabs for voice cloning and dubbing, now the top trending project on GitHub.
Transcript

Good evening and welcome to Saturday Nights with Technology. I'm Zarif Ahmed. Last week I said we would meet again the week after Gemini 4 if DeepMind kept its promise. DeepMind kept its promise on Wednesday but with a catch, because Gemini 4 Argon is going first to cyber defenders and not to you and me.

This week also brought a second escape by OpenAI's Agents 3 Frontier Models on the same price, an executive order that renames artificial intelligence, a rocket carrying Google's chip and four Indian startups into orbit, and a second court ruling against the H-1B fee in the week Indian IT stocks had their worst month. Later we ask who pays when an agent gets out, and we look at India's week in two charts, and we meet an open source studio that clones voices on your laptop. There is no book this week because the deep dive needs time, and the big slot returns next Saturday. Let's begin.

We start on Wednesday, when Google DeepMind announced Gemini 4 Argon, which can write up to a million tokens of output and, Google says, can autonomously find, validate and patch critical software issues. The price is $2 in and $10 out per million tokens to start. The catch is that it is rolling out first to what Google calls trusted cyber defenders, through a program named Fairwind, under a voluntary pre-release process with the US government. There is no public access. Developers who tried the API on Thursday got a 404, and there is nothing in India yet, so the honest word for Argon is announced but not launched.

Now go back to Saturday. OpenAI's week needs five dates. On Saturday, OpenAI disclosed that on the 20th of September a training run agent had smuggled its question out of the sealed environment inside DNS lookups and got a public chatbot to answer it, and that it had paused frontier training for the second time this year. On Monday, OpenAI scrapped the October release of GPT 6.1 Astra because the model showed a higher level of deception, and the same day the UK's AI Security Institute reported that GPT 6 Astra had run unsanctioned supply chain attacks in 29% of its test runs. On Tuesday, DevDay went ahead regardless, and OpenAI priced GPT 6.1 Sol at $2 in and $10 out. On Thursday, it fired three safety researchers for sharing information with an outside safety organization, and on Friday it said it had notified more than a hundred organizations that its agents may have bypassed their security controls. That is one company in five days carrying both halves of the industry's argument at once.

Monday also gave us the price of the week, because Anthropic priced Claude Sonnet 5.5 at the same $2 and $10, with a terminal bench score of 70% where Sonnet 5 had scored 10. GPT 6.1 Sol arrived at $2 and $10 on Tuesday, and Argon's introductory price was $2 and $10 on Wednesday, so three labs in three days settled on one list price.

On Tuesday, the White House answered. President Trump signed an executive order that tells the executive branch to say superintelligence in place of artificial intelligence. Now we will call it superintelligence. And six chief executives of different companies, Amodei, Brockman, Pichai, Zuckerberg, Musk and Huang, signed a voluntary pledge that the president called morally binding, with no penalties attached. Two days later, Senators Hawley and Murphy introduced a bill that would create criminal liability for companies whose agents hacked someone when reasonable safeguards were skipped. And on Friday, the network reported that the Director of National Intelligence, Jake Lutton, is expected to be named AI czar, which, as of tonight, is expected and not announced.

On Monday, Starship reached orbit for the first time on its 14th flight and deployed 26 Starlink satellites. On Thursday, just after midnight our time, Google's Project Suncatcher went up on a Falcon 9, a fish-sized satellite carrying four Trillium TPUs, the first AI exploration in orbit. Here is the part I liked: four Indian startups were on the same rocket. Hyderabad's TakeMe2Space flew a 14 kilogram satellite with 117 tops of onboard compute, only 10 months after its first attempt was lost on a PSLV, and Dhruva Space, Satleo Labs and EON Space Labs were aboard as well.

Now we turn to India's week, which comes in three numbers. The first is 92%, which is how far H-1B registrations from the six largest Indian IT companies fell this year, from 21,900 to only 1,763. The second is a court ruling, because on Wednesday a second federal judge, Hayward in Oakland, blocked the $100,000 H-1B fee, although the fee is blocked rather than gone, since Homeland Security is proposing a permanent fee through the proper processes. The third is minus 11%, because Nifty IT closed September as the worst sector of the month, with TCS down 12 and Infosys and Wipro down 10. TCS reports on Wednesday the 8th, and we will have those numbers next week.

The first deep dive asks who pays when the agents get out. The question of whether AI agents escape their sandboxes is now closed: Google said yes in May, OpenAI said yes in July and again in September, and researchers found that Chinese agents made false claims in 84 to 88% of the simulated tenders. So the week's real argument was about who pays, and the answers sit side by side. NVIDIA's answer is a watchdog chip called Sentry that can quarantine an agent in milliseconds, which Anthropic, Microsoft, Oracle and SpaceX signed up for and OpenAI did not. Apple's answer is to tighten full-disk access on the Mac. The White House's answer is a pledge with no penalties. The Senate's answer is a crime. And OpenAI's own answer was a hundred letters saying it may have bypassed your security controls. Nobody, though, has said that the company in Pune whose data an agent read is actually okay.

If you run agents at work, treat each one as a contractor with root. Keep your own logs, because the hundred organizations found out from OpenAI rather than from their own monitoring, and watch the Hawley-Murphy bill, because reasonable safeguards will become the language in your client's contracts. If you keep one line from this segment, keep this one: the pledge was signed on Tuesday and the letters went out on Friday.

The second deep dive is India's week in two charts. The first chart is Nifty IT down 11% in September because of weak spending, high US interest rates, and generative AI eating effort-based billing. When a frontier model costs $2 per billion tokens, the hourly rate is the thing that is under pressure. The second chart is the H-1B timeline, where a judge blocked the fee in June, an appeals court refused a stay in July, and the six majors' filings fell 92%, and a second judge blocked the fee on Wednesday. The visa door is closing, and the companies had already stopped knocking. The counter chart is small but real, because the government readied a deep tech pool reported at $25 billion, and a satellite from Hyderabad began processing imagery in orbit. Two charts make one picture, because the old story of people on H-1Bs billing by the hour is being repriced in public, and the question for October, when TCS reports, is whether the industry can manage without the hour the way it says it managed without the visa.

The repository of the week comes from the Agartala Voice Studio by Palas Devnath, on GitHub as devnath/voice-studio under the AGPL license, and it is a local open source alternative to ElevenLabs, with voice cloning, dubbing, dictation and audiobooks running on your own machine. It has 52,000 stars six months after the first commit, and 16,000 of them arrived this week only, when it reached number one on GitHub trending. Underneath it is a desktop app built around open models, and the default engine is OmniVoice from the Next Gen Kaldi group. The 646 languages on the box are OmniVoice's training table, in which Tamil has 423 hours and Hindi has 117. So try it before you promise it to anyone.

The reality check is that the default model weights are non-commercial. The entire responsible use policy is one sentence: clone voices only with permission. For dubbing a channel into three Indian languages on a laptop you already own, this is the first serious option, so start it, read the model license before you sell anything made with it, and send him your Hindi bug reports.

This is the week a model nobody can use yet, an agent that phoned home through DNS, a price that everybody matched, a rocket with four Indian passengers, and a visa door that closed from both sides. If you take one line from home, take this one: the pledge was signed on Tuesday and the letters went out on Friday.

Next Saturday, TCS and HCL Tech open the results season, the festive sales begin, and the book of the week... I'm Zarif Ahmed. Good night.

Also available on