BGAD Consulting
STRATEGIES. DELIVERED.
BGAD News Flash
Daily briefing: digital, tech and AI
25 September 2026
Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol and GPT-6 Luna within hours of each other. Opus 5.5 took the early acclaim on capability, while the two GPT-6 variants were positioned at the affordable end, pushing the frontier on price rather than on raw benchmark rank. The brief reads this as competition moving away from a single winner and towards cost per task and where each model sits in a working stack. For buyers the question stops being which model is best and becomes which model is worth running for which job.
Source: The AI Daily Brief, 23 September 2026, Opus 5.5 vs GPT-6 Sol and Luna
Anthropic says Claude, working in its biology lab, identified a previously undescribed enzyme system with characteristics resembling CRISPR, and did it in about 21 hours. The company has been open that it cannot yet say what the system is capable of, which makes this a preliminary result rather than a validated discovery. It is still the most concrete claim so far that a frontier model can surface genuinely novel biology rather than summarise existing literature. The unresolved part, what the system actually does, is also the part that matters for biosecurity.
Source: The AI Daily Brief, 24 September 2026, AI Agents Are Moving Into the Real World
Meta is bringing its Muse agent to smart glasses and to a new wearable, moving the agent off the phone and into always on hardware. That changes the interaction model from opening an app to speaking to something already switched on, which is the form factor bet Meta has been building towards for years. Sceptics continue to question whether consumers actually want an agent at all, but early users report real value in handing over the small administrative tasks nobody wants.
Source: The AI Daily Brief, 24 September 2026, AI Agents Are Moving Into the Real World
GrokBot is being rolled into Teslas, making the car a voice interface for a personal agent rather than a vehicle with a media system attached. Coming in the same week as Muse on glasses, it marks agents spreading into the devices people already sit inside for hours a day. The commercial logic is distribution. Whoever owns the surface people talk to without thinking about it does not have to win on model quality.
Source: The AI Daily Brief, 24 September 2026, AI Agents Are Moving Into the Real World
Kara Swisher and Scott Galloway work through whether Anthropic can credibly come to market at a two trillion dollar valuation, a number that would put a company with a few years of revenue history alongside the largest listed businesses in the world. The argument in favour rests on enterprise adoption and the scarcity of frontier labs available to public investors. The argument against is that the multiple prices in an outcome nobody can yet underwrite. Either way, a listing at that size would drag every private AI valuation onto a public mark.
Source: Pivot, 22 September 2026, Trump's Press Crackdown, Paramount Settles, and Anthropic's $2 Trillion IPO
The administration has rebranded its AI push under a Super Intelligence banner, a naming choice that signals ambition rather than caution and sits awkwardly against the safety framing used by the labs themselves. At the same time the United Nations General Assembly produced competing visions for how AI should be governed internationally, with no convergence on who sets the rules. The gap between a national industrial programme and a multilateral rulebook is widening, not closing. For anyone operating across jurisdictions that means planning for divergence rather than a single standard.
Source: The AI Daily Brief, 24 September 2026, AI Agents Are Moving Into the Real World
Eric Nguyen, chief executive of Radical Numerics, argues that the real biosecurity question is whether defensive capability can keep pace with what frontier models can already do in biology. His own work built genome language models, Evo and Evo 2, and used them to generate functional bacteriophage genomes, which is to say sequences that work rather than sequences that merely look plausible. The uncomfortable point is that the same capability curve serves both attack and defence, and only one side of it is organised and funded.
Source: Latent Space, 23 September 2026, Bio-security is an AI Arms Race, Eric Nguyen
Nguyen describes context windows for DNA models stretching from around 60,000 bases to three billion, which is the scale of an entire mammalian genome in a single pass. He also describes applying chain of thought style reasoning to sequence generation, with the model showing progressive optimisation when trained on ranked biological sequences. The next step is multimodal biology, extending beyond DNA to RNA, proteins, three dimensional structure, epigenetics and natural language in one model. Biology is being treated as a language problem, and the tooling is arriving faster than the governance.
Source: Latent Space, 23 September 2026, Bio-security is an AI Arms Race, Eric Nguyen
Frank Hutter, founder of Prior Labs, describes TabPFN, a transformer trained entirely on synthetic data that learns a full prediction algorithm in a single forward pass. He claims it outperforms the gradient boosted workhorses, XGBoost and CatBoost, on tabular machine learning tasks. His argument for why general purpose language models fail here is structural rather than a matter of scale. Large tables blow past token limits, and a language model has no native understanding of row and column invariance, so a purpose built architecture wins.
Source: Machine Learning Street Talk, 23 September 2026, How Deep Learning Finally Cracked Messy Tables, Frank Hutter
Alongside the model, Hutter introduces TabArena, an open and continuously updated benchmark for tabular algorithms using an ELO style score. Tabular data is where most enterprise value actually sits, and it has never had the public scoreboard that language and vision models have had for a decade. Hutter also discusses wiring TabPFN into coding agents through MCP servers, so an agent can run an end to end data science workflow rather than write code for a human to run.
Source: Machine Learning Street Talk, 23 September 2026, How Deep Learning Finally Cracked Messy Tables, Frank Hutter
Garrett Berntsen, chief AI officer at Accenture Federal Services and formerly deputy chief digital and AI officer at the Department of Defense, argues that APIs alone are not enough for AI to matter in national security. He draws on the U-2 programme, where the win came from how the thing was procured and how industry and government worked together, not from the aircraft. His prescription is institutional redesign, changing workflows, incentives and decision rights, and getting modern tools to front line service members rather than concentrating them at headquarters. The same diagnosis applies to most large organisations buying AI right now.
Source: ChinaTalk, 22 September 2026, Can AI Beat the Pentagon's Bureaucracy?
Eddy Lazzarin, a general partner at a16z crypto, argues that the safety debate over weights speculative superintelligence risk and under weights the cost of delaying useful technology. His alternative to new AI specific regulation is to use the tools that already exist, namely cybersecurity practice, legal liability, market incentives and stronger technical controls, and he floats the idea of models accumulating measurable trustworthiness reputations. He also warns against concentrating oversight authority in a small number of evaluators. Read the position with its interests visible. a16z holds extensive AI startup positions, and this is the firm's own partner arguing for the deployment pace those positions depend on.
Source: the a16z Podcast, 24 September 2026, The Case Against an AI Pause, Eddy Lazzarin
Ed Elson closes the episode on an AI data centre provider postponing its planned initial public offering. The company was not named in the episode, so treat the specific issuer as unconfirmed. The signal is what matters. Data centre operators have been the cleanest way for public investors to buy AI capital expenditure, and a withdrawn listing in that category suggests the bid is thinner than the capacity announcements imply.
Source: Prof G Markets, 24 September 2026, Bonds Are Going Haywire Again, Howard Marks Explains Why