The pitch for agentic payments usually goes like this: software that negotiates, pays, reconciles and optimises cash while the finance team sleeps. A new report from UOB Global Economics & Markets Research is a reminder that a payment involves more than moving the money. Someone also has to answer for it.

The bank’s argument is that autonomous payments will be shaped less by how capable AI becomes and more by how well financial institutions govern it. That means clear accountability, solid security controls and systems that work across institutions.

The model UOB describes is “bounded autonomy”. AI agents can execute transactions, but only within explicit mandates and defined permissions, under continuous monitoring, and with decisions that can be audited afterwards.

Also Read: The next AI payments boom may happen in the back office

That sounds sensible. It is also the framing you would expect from an incumbent bank.

The technology is ahead of the rules

UOB’s starting point is uncontroversial. The tools needed for an agent to start and complete a payment on its own are developing quickly. The problems are elsewhere. Liability frameworks have not been settled. There is no common standard for identifying an agent. And interoperability across institutions, countries and settlement networks is still limited.

The report calls these institutional problems, not engineering ones. Solving them requires technology providers, banks, regulators and policymakers to coordinate. That framing is more important than it first looks. A startup can ship a smarter agent in a quarter. Getting a central bank, three correspondent banks and a card network to agree on who pays when an agent buys the wrong thing takes much longer.

Southeast Asia is already working through versions of this. The Monetary Authority of Singapore has given financial firms one year to prepare for new AI risk rules. Even in the region’s most advanced market, companies adopting agentic AI still have thin audit trails. Regional tech leaders are cautious too: they are learning to trust AI agents, but not with production. If they won’t let an agent push code to live systems without a human check, they are unlikely to let one move company money unchecked.

Where agents could actually help: the treasury

UOB’s most concrete example is corporate treasury, not consumer shopping. In a mature system, it says, agents could help treasurers manage liquidity and coordinate financial workflows while staying within limits set by their institutions.

That makes sense in Southeast Asia. A mid-sized company operating across Singapore, Indonesia, Thailand and Vietnam deals with several currencies, several domestic payment rails and settlement windows that rarely line up. Agents that can route payments, time currency conversions and sweep idle cash could help Southeast Asian firms untangle cross-border payment costs. That is where much of the margin is lost today.

Also Read: How the next-generation neobank should be built for the agentic economy?

The infrastructure is improving too. Singapore’s PayNow Gen 2 upgrade and the gradual linking of regional real-time payment systems mean there are more rails an agent could use. UOB’s broader point, though, is that more rails also means more places where an agent’s authority has to be checked.

Not just a banking project

The banks are not the only ones building this. Ant International has launched what it calls full-stack AI-native tools for payments, FX and treasury. Circle’s US$400 million acquisition of Tazapay shows stablecoin companies moving directly into cross-border business payments. Visa and Mastercard are developing their own frameworks for agent-initiated transactions. Each group is effectively proposing a different answer to who should be trusted to oversee AI agents that spend money.

This is where some scepticism is useful. A bank arguing that governance, accountability and controls will decide the agentic payments race is arguing that the race will be decided by what banks already do well. That doesn’t make the argument wrong. It does make it convenient. The risk is that “bounded autonomy” turns into a reason to keep agents limited to tasks that don’t threaten existing fee income.

The building blocks are known. The timelines are not.

To be fair, UOB does name the components. It lists bounded mandates, portable identity frameworks, continuous authorisation models, cryptographic verification and trusted interoperability frameworks. It adds that stronger audit trails, human oversight and cooperation across the industry are essential. The goal, it says, is for an institution to be able to trace what an agent did, why it was allowed to do it, and who is responsible.

The identity question is the most urgent of these. As e27 has reported, payment risk is moving from KYC to KYA, from knowing your customer to knowing your agent. Today an agent typically acts with credentials borrowed from a human or a company. Without a portable, verifiable identity for the agent itself, every payment it makes depends on a credential that was never designed for that purpose.

What the report doesn’t provide, at least in the version made public, is figures. It has no adoption data, no estimate of losses caused by agents, and no target dates for when liability frameworks might be agreed. It describes a set of principles, not a delivery plan.

Trust is the product

UOB’s conclusion is its strongest point. The institutions that succeed, it argues, will not necessarily be the ones with the most autonomous systems. They will be the ones that make AI-driven financial activity trustworthy, controllable and accountable. If that happens, bounded autonomous finance could become a basic part of the future financial system.

Also Read: The end of manual finance? AI agents are coming for startup payments

For Southeast Asian fintech founders, there is a practical lesson. Selling an agent on how independent it is will not win over a corporate treasurer. Selling it on control is more likely to work: show exactly what the agent can do, what it cannot do, and produce an audit log a regulator would accept. The startups that do well here may be the ones that build those controls well, not the ones with the most impressive demos.

The post Banks want AI agents to move money, as long as someone is accountable appeared first on e27.

On 14 October 2026, e27 hosted Borderless Innovation: X-HUB Tokyo-Singapore Demo Day 2026 with X-HUB TOKYO, the Tokyo Metropolitan Government’s global startup accelerator, run by Deloitte Tohmatsu. This is the third year e27 has partnered with Deloitte Tohmatsu to bring Tokyo’s startups to Southeast Asia, and this year’s cohort arrived with a clear focus: technology built for Japan’s labour shortage, ready for a region starting to feel the same pressure.

The online session brought nine Tokyo-based startups in industrial AI, robotics, enterprise AI, supply chain, medical AI and water technology in front of Singapore and Southeast Asian investors, corporates and government agencies. Each founder pitched, took live questions from a mentor on Southeast Asia market fit, and then met guests in themed breakout rooms. The session came two weeks before the cohort travels to Singapore for SWITCH 2026, from 27 to 29 October.

Industrial AI and robotics tackling the labour shortage

• Hutzper: builds Owlox, a safety AI for construction sites that checks whether workers are actually clipped in, not just wearing a harness.
• SYNRA: uses cameras to read tiny vibrations in equipment and predict failures, with no sensors needed on machines that are too hot or too high to reach.
• D’isum: brings predictive maintenance to SME manufacturers through a smartphone or tablet, with no training data or data science team required.
• RoboFounders: builds physical AI and a robot OS for dangerous manufacturing work, with deployment results already in Malaysia.

AI agents built for enterprise operations

• Jurin AI: runs AI agents that handle customer support across phone, email and chat in more than 105 languages, around the clock.
• datagusto: provides a runtime safety layer that catches AI agents before they misfire, then diagnoses and fixes the root cause.
• AI Hikari: runs supply chain work, from product listing to delivery and settlement, through AI agents on its COMSEE platform.

Health and food security solutions with regional reach

• Neuspective: helps radiologists and pathologists produce faster, more accurate diagnostic reports with its Qreia platform.
• Shelltas: runs a “Water OS” that keeps aquaculture and hydroponic systems stable with far fewer water changes.

Also Read: Why Antler is going all-in on Japan’s earliest-stage founders

A format built for conversations, not just pitches

Each startup pitched for three minutes, then took two minutes of live questions from a mentor. The mentors focused on one thing: whether the product fits Southeast Asia, and who the first partners and customers could be.

The five mentors were Remi Choong, Partner at Elev8.vc; Andrew Shuttleworth, Chief Strategy Officer at Obayashi Construction-Tech Lab Singapore; Syahmi Rashid, Technology Scout and Solutions Strategy Consultant at the Infocomm Media Development Authority (IMDA); Charmaine Ng of Nebius’s startup platform and community team; and Tushar Poddar, Advisor and Principal Consultant at TPM Solutions.

After the pitches, Choong and Shuttleworth joined Christine Galolo, e27’s Director for Growth and Marketing, for a fireside chat titled “Why Southeast Asia, Why Now: Japanese Deep Tech Going Global.” The pairing was deliberate: a Japanese corporate already operating in Singapore, and a Singapore deep tech investor.

The session closed with four themed breakout rooms: Smart Factory, Robotics and Infrastructure, AI Agents, and Health and Food Security. Guests picked the room they wanted when they registered, and we matched anyone who didn’t. Each room had an e27 facilitator and a mentor alongside the founders.

What the demand told us

The event drew 141 registrations and 96 live participants. The external guests came from 59 organisations, a mix of venture capital firms, global corporates, government agencies and development finance institutions.

In the days after the session, attendees sent 46 requests to connect with specific founders, and every one of the nine startups received interest. The room was buyers and capital. Nearly two in three requests came from investors and corporates, not ecosystem observers. That matters for a cohort heading into in-person meetings.

Some of those connections are already turning into meetings during SWITCH week.

Also Read: Japan’s next-gen startups land in Singapore at X-HUB TOKYO 2025

What this means for Japan-Southeast Asia collaboration

Japanese startups have long been seen as strong on technology and slow on going abroad. This cohort suggests that view is out of date. These founders came with clear Southeast Asia use cases, live deployments at home, and in some cases results already in the region.

The appetite on the Southeast Asian side is just as clear. The technology Japan built for its own demographic challenge is exactly what the region’s factories, construction sites and service businesses are starting to need.

The bridge between the two works best when it is built before the plane lands. Demo days like this one are where that bridge starts.

Thank you to Deloitte Tohmatsu, X-HUB TOKYO and the Tokyo Metropolitan Government, to our mentors and fireside speakers, and to the nine founding teams who pitched.

—

This article is produced by the e27 team

We can share your story at e27 too! Engage the Southeast Asian tech ecosystem by bringing your story to the world. Reach out to us here to get started.

The post What 9 Tokyo Deep Tech Startups Learned About Southeast Asia’s Appetite for Japanese Innovation appeared first on e27.

Betagro has taken majority control of Freshket, the Bangkok platform that supplies ingredients to hotels, restaurants and caterers, through its venture arm Betagro Ventures. Neither side disclosed the price, the stake size or what happens to Freshket’s venture backers.

This is less a courtship than a marriage after a long engagement: Betagro Holding, the listed group’s largest shareholder, joined Freshket’s US$23.5M Series B in 2022 alongside OR and Openspace Ventures. A later round brought in Thai President Foods, maker of MAMA noodles.

Founded in 2017, Freshket lists more than 10,000 products for over 20,000 foodservice operators across Bangkok, Pattaya, Hua Hin and Khon Kaen. The new capital will fund cold-chain upgrades, a wider imported range and a push into more regional cities.

The deal fits a pattern. As venture money thins, Thailand’s corporates are deciding which startups survive and on what terms, and in food and farming the realistic finish line is often a strategic sale rather than a listing. The awkward question is neutrality. Freshket sold itself as a neutral bridge between farmers and kitchens; its new owner produces the meat, eggs and processed foods on its shelves. Restaurants will watch whether Betagro products start ranking higher.

Editor.

—

Regional

Salmon raises US$140M to push motorbike and MSME loans: The Manila-based lender’s second raise in six months lifts disclosed financing to US$378M. It has not revealed the equity-debt split, the lead investor or a valuation, and still withholds loan-book and bad-loan figures.

VinFast revenue jumps 85% in Q2 as losses keep widening: Revenue reached US$1.2B on 70,085 EV deliveries, up 96%, while its continuing operations loss rose 128% to US$1.2B. Markets outside Vietnam took only about 11% of deliveries, keeping the export story thin.

Aolani lines up 500MW of AI data centre capacity in Malaysia: The Singapore-founded neocloud expects capacity from the second half of 2027 and projects up to 580MW by 2027, up from roughly 80MW this year. The letters of intent are non-binding.

Singtel and Ericsson sign five-year pact on 5G Advanced and AI: A live trial of AI-powered radio access delivered a 19% throughput gain. Singtel plans network slicing for autonomous vehicles, robotics and logistics, eyeing Changi Airport Terminal 5 and new JTC industrial sites.

BEENEXT and Lotus back Miracly’s US$2M bet on AI brand visibility: Swedish-founded The Wild Ventures launched the platform to track how chatbots describe brands. With Profound reportedly valued at US$1.8B, mid-market Asian brands may be Miracly’s better bet.

Three in four Singaporeans think their spare cash underperforms: A Milieu survey of 1,000 residents found 89% hold spare cash, yet many have little idea what their savings accounts pay, even as they chase cashback and rewards points.

Singapore staff say promised burnout support is nowhere to be seen: Jobstreet by SEEK’s 2026 workplace culture study finds half have never seen the support employers claim to offer, exposing a gap between HR handbooks and daily reality.

Interviews & Features

Ex-Bukalapak engineer builds legal AI while fighting graft case: Ibrahim Arief, a defendant in Indonesia’s Chromebook procurement case, built Vindica for defence teams facing live trials. The US-incorporated startup has a paying client and angel backing, while his five-year sentence awaits the Supreme Court.

AI exits are booming, but SEA startups are far from the action: Amazon is reportedly in talks to pay about US$7B for Decart, valued just under US$4B in May. Southeast Asia remains far removed from this exit frenzy, the analysis argues.

Pine Labs bets Singapore can anchor its unified payments push: The Indian fintech combines acquiring, instalments and engagement on one platform. Its regional business chief argues merchants want fewer providers, simpler reconciliation and a single integration as payments turn real time.

14 SEA gaming and sports-tech startups that outlived play-to-earn: After the 2021 Web3 gaming bust, the region’s best play businesses held on, and many never raised much venture money. The roundup maps who is still standing.

International

Anthropic and Nscale reportedly near a second US$45B compute deal: Bloomberg says terms could still shift. A deal would follow August’s six-year US$45B agreement for West Virginia capacity, as both firms chase IPOs and Nscale faces pushback over a US$30B valuation.

Flipkart tests food delivery to break the Zomato-Swiggy duopoly: The Walmart-owned marketplace has expanded trials to 8,000 employees and pitched restaurants a 10% base commission, against the 20-25% incumbents charge. A public launch is expected by end-October.

Replit’s Amjad Masad pitches the ‘self-driving company’: At TechCrunch Disrupt, Masad said agents now handle grunt work at Replit, leaving staff to decide what to build. Humans, he argued, will be valued for judgement and taste.

Cybersecurity

India says rogue AI agents breached firms’ live systems undetected: CERT-In said AI models under test read confidential business records; providers, not victims, flagged every incident. It blamed exposed credentials and over-permissioned cloud setups, urging sandboxed agents and human sign-off.

Horizon3 picks Singapore for ASEAN HQ with EDBI backing: The San Francisco security-testing firm, valued above US$2B in August, will also serve Australia and Japan from Singapore. SG Growth Capital co-invested; its NodeZero platform probes live networks continuously.

Half of ASEAN scam cases go unresolved as AI fraud spreads: The GSMA’s ASEAN Consumer Scam Report 2026 found nearly nine in ten consumers had spotted AI in scams, from deepfakes to voice cloning, a trust test for telcos embedding AI in apps.

Banning AI won’t stop staff pasting contracts into chatbots: Most firms in Singapore and Thailand already have AI policies, yet employees still use unsanctioned tools. Kazu Fukuda sets out what companies should do instead of bans that push usage further from view.

Semiconductor

Malaysia unveils six homegrown chips as it pivots to design: The designs carry combined sales potential of US$73.5M, but the economy minister conceded that winning customers and passing industry qualification lie ahead, the real test of Malaysia’s push to own chip IP.

China moves to curb exports of 23 AI technologies: The draft, open for comment until 15 November, reportedly covers some AI chip-design technologies and requires exporters to register transfers now. API and SaaS access to Chinese models largely escapes the rules.

BMI says AI capex will keep driving APAC chip demand into 2027: Taiwan, South Korea and Singapore benefit most, while Malaysia draws more data centre investment. BMI warns a hyperscaler slowdown or fresh trade restrictions could hit semiconductor exports.

AI

GoTo releases 27B open-weight model for Indonesian languages: Built on Gemma 3 with AI Singapore, it supports Javanese, Sundanese and Balinese alongside Indonesian, English and Batak Toba. Running it at full precision needs about 54-60GB of video memory.

Vertex leads RapidCanvas’s US$29M Series B in enterprise AI: The Austin-based firm, now with US$52.5M in funding, will expand field AI engineering in the US, Brazil and India. Hitachi Ventures joined; RapidCanvas reports eightfold ARR growth over six quarters.

UOB says AI agents can move money, but only on a leash: Its research arm backs bounded autonomy with auditable mandates, citing unsettled liability rules and no common agent identity standard. The report offers principles but no adoption data, loss estimates or timelines.

HSBC’s AI jobs optimism needs a closer look in Southeast Asia: Entrepreneurs across 17 markets plan US$367B in AI spending, HSBC’s Global Entrepreneurial Wealth Report 2026 finds. The analysis argues who is spending matters more than the headline figure.

Singapore’s Gen Z is bringing its own AI to work: A Deel study with YouGov and ThinkerTank shows young Singapore workers using AI on personal devices, often without oversight. SMEs, the piece argues, need clear usage rules rather than outright bans.

Thought Leadership

As VC retreats, Thai corporate investors call the shots: Yvan Goudard maps how bank-linked vehicles such as SCB 10X, Beacon Venture Capital and Bualuang Ventures gain leverage over Thai startups as independent venture money thins.

When AI becomes the gatekeeper, brands must win machines first: Alon Weinstock argues that after search, social media and algorithms, AI now sits between brands and buyers, so marketers must persuade the filter before reaching the human who still decides.

SEA’s digital leap has made its networks faster, and more fragile: Hans Osnabrugge argues that as fibre, smartphones and online public services replaced older infrastructure, payments, commerce and government now depend on networks that need far stronger resilience planning.

Can Bitcoin hold US$80,615 or face a liquidation cascade?: Anndy Lian reads Bitcoin’s 0.81% dip to US$81,556 through shifting global liquidity, noting an 85.9% correlation with gold, a sign the asset now trades on macro forces.

Founders need friction, until one thorn goes too deep: Will Fan argues tough investors, demanding customers and abrasive high performers force decisions comfortable companies avoid, but leaders must spot when that friction turns damaging.

How one small B2B team uses AI to triage inbound leads: Automaid founder Daniel Tan uses AI to summarise leads and draft replies, but a human approves every response. The aim is a consistent, context-aware first reply, not more automated outreach.

Using AI to keep cross-border partnership details straight: Stella Kim explains how she uses AI to draft follow-ups after overseas partner meetings, capturing what each side agreed and what still needs clarifying.

What happens when AI takes your side in a workplace row: Dr Zoe Wyatt-Potage warns that workplace conflicts rest on interpretation, and chatbots that validate our version can harden assumptions about colleagues instead of testing them.

The post Ecosystem Roundup: Betagro buys control of Freshket as Thai corporates close in appeared first on e27.

For most of the generative AI boom, someone in North Sumatra typing a question in Batak Toba has been talking to a machine that, at best, half understands them. GoTo Group says it wants to change that. But the hard part is only beginning.

Indonesia’s largest digital ecosystem on Monday released a 27-billion-parameter language model trained for Bahasa Indonesia and a clutch of regional languages, unveiling it at Bali Tech Week and publishing the weights on its official Hugging Face repository.

The instruction-tuned model sits on Google’s Gemma 3 architecture and handles six languages: English, Indonesian, Javanese, Sundanese, Batak Toba and Balinese. It also comes with a 128,000-token context window, which, in plain terms, means it can read a lengthy contract or a long customer-service conversation in one go rather than in chopped-up pieces.

Also Read: Google bets on schools, telcos and local languages to scale Gemini in Southeast Asia

The repository names GoTo as the initiator, and GoTo and AI Singapore as co-developers. That cross-border credit matters. It places the release inside a broader regional push, in which AI Singapore has been a recurring anchor, to fix a problem global models have never taken seriously: Southeast Asia’s languages are spoken by hundreds of millions of people, yet treated as an afterthought in most training data.

“Open source”, with an asterisk

GoTo calls the model open source. The model card tells a slightly different story: the governing licence is Google’s Gemma licence, not a standard open-source licence such as Apache 2.0 or MIT. Gemma’s terms permit broad use and modification, but they come with Google’s own usage restrictions.

For a hobbyist, the distinction is academic. For a bank, a telco or a government agency planning to embed the model into a product, it is the first thing the legal team will flag.

This is not a GoTo-specific sin. Much of what the industry markets as “open” AI is really “open weights”: you can download and tinker, but someone else still writes the rules.

Still, for a release explicitly framed around reducing dependence on foreign platforms, building on a licence controlled by a US tech giant is an irony worth noting. It sharpens a question regional commentators have been raising for months: whether Southeast Asia can build sovereign AI on borrowed choices.

From Sahabat-AI to a leaner bet

The release is not GoTo’s first foray into homegrown language models. In 2024, the company and Indosat Ooredoo Hutchison launched the Sahabat-AI initiative, which later produced a 70-billion-parameter model and a multilingual chat service. GoTo positions the new model as part of that broader effort rather than a replacement for everything that came before.

The shift to a smaller 27-billion-parameter model looks pragmatic. Bigger is not always better when your target users are Indonesian universities and startups, not hyperscalers. The timing is also telling. GoTo only recently posted its first quarterly net profit after years of restructuring and subsidy cuts, and its leadership under CEO Hans Patuwo has every incentive to show investors that AI is a strategic asset, not just a cost centre.

On training data, GoTo says it drew on Indonesian-language text from media partners and public-sector websites, and worked with Universitas Sumatera Utara on Batak-language material. None of this has been independently audited. That gap matters beyond accuracy. As newsrooms globally fight over whether, and on what terms, their journalism should feed AI systems, the precise nature of GoTo’s arrangements with its media partners deserves more transparency than a single line in an announcement.

The 60GB question

Here is where the “AI for everyone” pitch meets the server room. According to the model card, running the full-precision model requires roughly 54GB to 60GB of video memory. The suggested hardware is one Nvidia A100 80GB card, two Nvidia L40S cards or three RTX 4090s.

Translation: this is not something a two-person startup in Yogyakarta runs on a gaming laptop. Most smaller organisations will need cloud infrastructure, specialised servers, or a compressed, lower-precision version that trades some quality for affordability. GoTo’s repository offers deployment guidance for popular frameworks such as vLLM and Hugging Face Transformers, which helps, but it does not lower the bill.

Also Read: Why Singapore, Indonesia, and Vietnam are losing the AI race they think they are winning

That makes Indonesia’s compute strategy part of the story. Jakarta has been courting Nvidia and AWS to build out domestic chip and cloud capacity. A local-language model is only as sovereign as the infrastructure it runs on, and if most deployments end up on foreign-owned clouds anyway, the independence argument weakens considerably.

Benchmarks without a referee

The model card lists Indonesian and Southeast Asian evaluation suites, including IndoMMLU and SEA-HELM. But GoTo has not published independently verified comparisons against rival models, so claims of superior local-language performance remain claims. The repository itself is refreshingly candid: it warns that the model can generate inaccurate or misleading information, may stumble on reasoning-heavy tasks, and should not be treated as authoritative without verification. Safety testing and safeguards, it adds, are the developer’s responsibility.

A crowded field of local champions

GoTo is entering a busy arena. AI Singapore’s own SEA-LION family has long targeted regional languages, Thailand’s SCB 10X has pushed its Typhoon models for Thai, and Malaysia has made local context a sovereign AI priority.

Above them all sit global giants (Google, Meta, Alibaba) whose general-purpose models keep improving in Southeast Asian languages, partly because regional users increasingly prompt them in Vietnamese, Thai and Indonesian. GoTo’s edge, if it has one, lies in lower-resource languages such as Batak Toba and Balinese, where global players have little commercial reason to invest.

What it adds up to

GoTo frames the release as infrastructure for Indonesia’s long-term digital agenda, a foundation universities, startups and public institutions can build on without shipping every workload abroad. The ambition is worth applauding, particularly for languages that the AI industry has historically ignored.

Also Read: Featherless AI secures backing to make open-source models viable at enterprise scale

But access to weights is not adoption. Teams will still need a licensing review, suitable hardware, clean application data and rigorous testing in the dialects and domains they actually serve. GoTo has opened the door. Whether Indonesia’s developers can afford to walk through it is the real test.

The post GoTo’s 27B model speaks Batak, but the real test is who can afford to run it appeared first on e27.

In 2015, Demis Hassabis looked back at a decision he had made more than a decade earlier. By 2004 he was already thinking about what would eventually become DeepMind, but concluded that the technology was not ready. As he later put it, “we still didn’t have enough of the components.” Deep learning had not yet taken off, computing power was inadequate and many of the systems needed to build something approaching general intelligence simply did not exist.

By 2026, Hassabis was describing a fuller set of ingredients: Gemini as a world model, AlphaGo-style search and planning, specialist tools, memory, reasoning and systems capable of acting in the world. The possibility is that Google has not recently changed its theory of artificial general intelligence. It may finally have enough of the pieces to execute a theory DeepMind has pursued for years.

That matters beyond the contest among AI labs. Falling AI costs make individual tasks cheaper to automate, but cheap task execution does not necessarily replace a worker. Jobs are bundles of tasks. Planning, memory, simulation, tool use and action let an AI system move across that bundle. If Google can combine them reliably, the shift from AI as a tool to AI as a worker could happen before anyone agrees that AGI has arrived.

DeepMind was never really building a chatbot

DeepMind’s early history looks very different from modern large language models. Its first major breakthrough was not language generation but an agent learning how to act. In 2013, DeepMind showed that a neural network could look at raw pixels from Atari games, decide what action to take, observe the result and improve from the reward. Hassabis later described the system as moving from “pixels to action”.

Planning became another strand of the same programme. AlphaGo combined neural networks with search, examining possible future positions before deciding what to do. In 2016 Hassabis said one purpose was to see how neural networks could be “meshed with planning approaches”. A year later DeepMind was experimenting with imagination-augmented agents that simulated possible futures before acting.

DeepMind was also working on memory and continual learning. The projects differed technically, but together they reveal a consistent view of intelligence: a capable system would need to perceive, remember, plan, imagine, act and learn over time. What these systems lacked was generality. AlphaGo could defeat the world’s best Go players while being useless at almost everything else.

Also Read: Does the iPhone show us which industries AI will destroy?

Then the LLM arrived

Modern large language models changed that part of the problem. Google Brain’s Transformer architecture, published in 2017, became the basis for systems trained across enormous quantities of language, code, images and other information. As they scaled, these models developed broad competence across subjects and tasks that earlier DeepMind agents could not approach.

Former DeepMind researcher Misha Laskin has described the distinction neatly: language models provide “generality”, while reinforcement learning provides “capability”. DeepMind had systems that could act effectively but only inside narrow worlds. LLMs were almost the reverse: remarkably broad, but unreliable when required to pursue objectives autonomously across long sequences of actions.

In 2023 Google combined the two research traditions institutionally when it merged Google Brain and DeepMind into Google DeepMind. Gemini appeared soon afterwards. From the beginning, Google described it as multimodal, designed to work with tools and intended eventually to support memory and planning. Hassabis went further, describing Gemini as combining some strengths of AlphaGo-type systems with the language capabilities of large models.

That makes Gemini easier to understand as more than Google’s answer to GPT. The large language model may instead be the general-purpose cognitive core around which DeepMind’s older work on search, planning, memory and action can finally be assembled.

The old architecture is reappearing

Google’s recent projects increasingly resemble components DeepMind once had to investigate separately. Hassabis has said the ambition is a “world model, not just a language model”. Google’s Genie project explores that direction by generating interactive environments in which actions alter subsequent events. SIMA supplies the agent side: Google has combined Gemini with SIMA so an AI can interpret objectives, reason about them and act inside 3D environments.

Deep Think attacks another weakness of conventional language models by exploring several possible reasoning paths before committing to an answer. Robotics extends the same structure into the physical world. Gemini Robotics ER 2 uses Gemini 3.5 Flash as a high-level reasoning system that can interpret a situation, plan multiple steps and coordinate lower-level models that produce physical actions.

That modularity may be crucial. Human workers do not personally contain every capability they use. They understand an objective, choose a tool, interpret its output and decide what to do next. An artificial worker could operate similarly. This also explains Hassabis’s emphasis on specialist tools: the general model need not contain the best protein-folding system or robotics controller. It needs to know when to call one and what to do with the result.

Also Read: The AI boom won’t help you scale. Your unit economics will

The economic threshold is below AGI

The important economic threshold is much lower than full human-level intelligence. Consider an analyst. The job may require understanding an objective, finding information, deciding what matters, operating software, creating an output, checking it, recognising mistakes, revising the work and deciding when it is finished. An AI system that can draft one paragraph cheaply automates a task. One that can perform most of that sequence reliably begins to compete with the role.

That is why falling cost and expanding capability matter together. Inference has been getting cheaper. At the same time, search, planning, world models, tools and memory expand the number of adjacent tasks AI can perform. More of the job therefore becomes technically automatable just as the price of automating it falls.

Better reasoning often requires more computation: several candidate solutions, verification and repeated tool calls. If inference remains expensive, those techniques can make autonomous AI uneconomic. As inference prices fall, more computation can be spent making the system reliable while the total cost of completing the work still declines. Different occupations can therefore cross the threshold at different times, before any universal AGI moment.

Google may still fail

None of this means Google has solved the problem. World models can be wrong. Search can amplify errors. Reinforcement learning works best where success can be measured clearly, whereas many real jobs involve ambiguous objectives and human judgement. Long autonomous workflows accumulate errors, and persistent memory can preserve bad information as easily as useful experience.

Google itself is also showing signs of execution difficulty. Gemini 3.5 Pro missed its expected release window. In August Hassabis handed day-to-day leadership of Google DeepMind to Koray Kavukcuoglu and became Alphabet’s Chief Scientist. Google says the change allows Hassabis to concentrate more heavily on long-term AGI research. A less favourable interpretation is that the long-horizon research programme and commercial Gemini roadmap are becoming more separate.

The architecture may never converge in the way the historical pattern suggests. Genie, SIMA, Deep Think, robotics and memory could remain distinct research programmes while Gemini itself improves mainly through better training and more compute. If AI becomes cheaper and better at individual tasks but remains unreliable when asked to complete long workflows without supervision, the thesis will have failed its clearest test.

But the history makes Google’s current direction more credible than recent announcements alone would suggest. DeepMind was pursuing reinforcement learning, memory, planning, simulated futures and action before modern LLMs existed. Those systems lacked broad generality. Transformer-based models supplied much of that missing capability. Google then merged the two research traditions and began building world models, agents, reasoning systems and robotics around Gemini.

That does not prove Google has discovered the route to AGI. It does suggest that its current work is not merely a collection of fashionable additions to a chatbot. Google may be assembling the machine DeepMind wanted to build before the components existed. If it works, the economically important date will probably not be the day Google announces AGI. It will be the quieter point when an AI system becomes cheap, reliable and broadly capable enough to do the whole job.

For more information on this proprietary research framework analysing AI-driven economic change, please visit here. 

—

Editor’s note: e27 aims to foster thought leadership by publishing views from the community. You can also share your perspective by submitting an article, video, podcast, or infographic.

The views expressed in this article are those of the author and do not necessarily reflect the official policy or position of e27.

Join us on WhatsApp, Instagram, Facebook, X, and LinkedIn to stay connected.

The post Google may already be showing us its route to AGI appeared first on e27.