BGAD Consulting
STRATEGIES. DELIVERED.
BGAD News Flash
Daily briefing: digital, tech and AI
5 October 2026
Google has started putting internal numbers behind its new frontier model. The claims relayed on the show include more than 300 TiB of data centre memory freed and code migrations running to 800,000 lines, with Sundar Pichai saying teams already use the model extensively in house. Against that, Bloomberg reports that some Google employees with access say Argon struggles on certain coding tasks, which Google disputes. Logan Kilpatrick's counter is that new Gemini revisions are tested by thousands of engineers over weeks before release.
Source: The AI Daily Brief, 2 October 2026, Gemini 4 Argon, Sonnet 5.5 and What Matters with AI Models
Google DeepMind has split its robotics release into three pieces. Gemini Robotics ER 2 is the reasoning layer, built on the Gemini 3.5 Flash line and handling semantic understanding and planning. Gemini Robotics 2 is the vision language action model that actually executes, and a smaller on device model covers offline or low connectivity deployment. The genuinely new capability is whole body control, where the action model reasons over the entire machine from fingertips to feet rather than treating walking and manipulation as separate systems. DeepMind builds that work with Apptronik, Agile Robots and Boston Dynamics.
Source: The Cognitive Revolution, 3 October 2026, One Brain, Any Body: Google DeepMind's Keerthana on Gemini Robotics 2, Cross-Embodiment and Humanoids
Having delayed its public listing on safety grounds, OpenAI is going back to the private market instead. The raise discussed on Pivot is 30 billion US dollars at a valuation of 1.4 trillion. The hosts read the combination as a company that wants public market money without public market disclosure, at a moment when its safety record is the thing most likely to be scrutinised in a prospectus.
Source: Pivot, 2 October 2026, AI's Rocky Road to Wall Street, Hegseth's Macho Military, and Trump's AI Safety Theater
The detail that emerged after the signing is how the voluntary frontier safety accord is actually meant to bite. Internal controls are verified by independent teams, validated by external professional auditors, with audit reports going to independent board committees that carry a fiduciary duty to act. The FTC and SEC retain enforcement authority over the public commitments. David Sacks argued on air that the resulting governance is de facto mandatory and can be implemented immediately without new law. Worth noting that Sacks holds a US government AI policy role alongside running Craft Ventures, so he was defending an agreement he helped produce.
Source: All-In, 2 October 2026, Trump's Super Intelligence Summit, AI Safety Accord, GDP Beats, Midterm Predictions
The early read on Anthropic's new mid tier model is that it is unusually token hungry. At maximum settings it runs about 7.60 US dollars per task, and its maximum setting test run on the Artificial Analysis index cost more than Opus 5.5 did. It scores 56 on that index, second only to Opus 5.5. The practical conclusion drawn on the show is that Sonnet 5.5 works best as a subagent under Opus rather than as a standalone cheap option, which is the opposite of how a mid tier model is usually positioned.
Source: The AI Daily Brief, 2 October 2026, Gemini 4 Argon, Sonnet 5.5 and What Matters with AI Models
The most interesting result discussed on Latent Space is about harness design rather than model scale. Recursive language model systems trained on shorter tasks showed eight to thirty times generalisation to longer ones, which suggests the harness is doing work people have been attributing to the weights. Systems built on the same approach produced early solutions on ARC-AGI-3 before OpenAI's Astra did, and Harvey has post trained one on document analysis. The underlying thesis is a capability overhang created by wrapping increasingly strong models in primitive scaffolding.
Source: Latent Space, 2 October 2026, Academia is for Ambition, Alex Zhang, MIT
Keerthana Gopalakrishnan, who leads Gemini Robotics research at DeepMind, was asked to rate robotics on a one to six scale of how many GPTs it has travelled. She said GPT-2, the same answer she gave a year ago. Her reason is cross embodiment: a policy that only works on one robot body is not yet a general brain the way a language model behaves identically on any computer. The honest read is that the demos are improving faster than the underlying generality.
Source: The Cognitive Revolution, 3 October 2026, One Brain, Any Body: Google DeepMind's Keerthana on Gemini Robotics 2, Cross-Embodiment and Humanoids
The structure of the filing is itself the story. Anthropic's S-1 devotes roughly 80 pages to risk factors against 48 pages describing the business, and the risk section explicitly warns about models that can resist shutdown and manipulate information. The filing seeks a valuation near two trillion dollars and discloses a 42 billion dollar net loss for 2025, of which about 34 billion is accounting charges rather than cash burn. A company arguing its product may be dangerous while asking the market to price it as infrastructure is a new kind of document.
Source: Pivot, 2 October 2026, AI's Rocky Road to Wall Street, Hegseth's Macho Military, and Trump's AI Safety Theater
Meta's personal agent is growing faster than any comparable launch cited on the show. Muse went from about 500,000 weekly users at the end of its first week to three million weekly and one million daily users in slightly over three weeks. For comparison, Codex took roughly three months to reach three million weekly users. Consumer distribution, not model quality, is the variable doing the work here.
Source: The AI Daily Brief, 2 October 2026, Gemini 4 Argon, Sonnet 5.5 and What Matters with AI Models
Alex Atallah of OpenRouter and Amjad Masad of Replit made the case that the future is an ecosystem of specialised models working together rather than one all purpose model. Atallah described routing differently trained models by task, with model fusion used to reach frontier level performance at lower cost, and argued smaller models are both cheaper and safer. Masad's companion point is that enterprises increasingly need independence across models, clouds and data. Read with the usual caution: this is an a16z podcast, both guests run companies whose economics depend on that thesis, and Replit is an a16z portfolio company.
Source: the a16z Podcast, 3 October 2026, Beyond the God Model, Alex Atallah and Amjad Masad
Gopalakrishnan described hardware moving faster than almost anything else in the stack, comparing two hands she has used directly. One has roughly the grip strength of a ten year old and is built on a decade old design. The newer, stronger alternative she has watched open jars. She also set out where simulation still fails: locomotion trains well in simulation because flat floor contact is easy to model, while cloth, friction and deformable objects are where the transfer from simulation to the real world still breaks down.
Source: The Cognitive Revolution, 3 October 2026, One Brain, Any Body: Google DeepMind's Keerthana on Gemini Robotics 2, Cross-Embodiment and Humanoids
Buried in the same filing is a channel concentration number that matters more than it looks. Just under half of Anthropic's sales reach customers through Amazon and Google, both of which are investors and both of which sell competing models. That is a different exposure from customer concentration: it means the two companies best placed to substitute their own product also control the shelf. Any pricing or bundling move by either one lands directly on Anthropic's revenue line.
Source: Pivot, 2 October 2026, AI's Rocky Road to Wall Street, Hegseth's Macho Military, and Trump's AI Safety Theater
Agentic commerce is splitting into companies that want the traffic and companies that do not. DoorDash has launched a text based ordering agent that matches a phone number to a customer profile, typical orders, delivery address and payment details. Agentic grocery orders are reported to carry roughly 50 percent higher basket values, though the show does not name a source for that figure. At the same time Amazon is blocking Muse agents from its properties, which tells you how the incumbent retailers see the same trend.
Source: The AI Daily Brief, 2 October 2026, Gemini 4 Argon, Sonnet 5.5 and What Matters with AI Models
Sacks put the constraint on deployment at the grid rather than the fab. His claim is that American electrical generation has been broadly flat for 25 years while China doubles its grid capacity roughly every decade, and he called opposition to data centre construction misguided on that basis. Treat the figures as host assertions rather than audited numbers. The investor caution applies here too, since the hosts hold positions across the infrastructure build out they are arguing for.
Source: All-In, 2 October 2026, Trump's Super Intelligence Summit, AI Safety Accord, GDP Beats, Midterm Predictions
Luca Ferrari, co-founder and chief executive of Bending Spoons, set out how the Milan based company has assembled AOL, Evernote, Vimeo, WeTransfer and Meetup. His argument is that the model differs from private equity in that the company operates rather than financialises, cutting headcount sharply after acquisition and then rebuilding the products with AI. It is one of the few examples of a European firm consolidating American consumer software rather than the other way round. Note that no revenue or acquisition figures were disclosed on the public episode page.
Source: Prof G Markets, 4 October 2026, This EU Firm Made Billions Buying Up Forgotten Tech