22 minutes
Adam Barty
In this quarters AI news wrap we discuss DeepSeek’s challenge to OpenAI with low-cost training claims, how big players like Telstra and CBA doubling down on AI, as well as looking at Apple’s AI stumbles. From security risks to emerging tools like OpenAI’s Operator, this quarter’s AI highlights reveal big shifts, big bets, and big questions.
DeepSeek has made waves in the AI industry, not primarily for its output quality, but for its efficient operation and cost-effectiveness. While matching competitors like OpenAI in performance, its significantly lower operating costs have attracted attention, including integration into Microsoft's API offerings. However, security concerns persist, with data breaches and storage issues in China leading to government restrictions. Meanwhile, other significant developments include Telstra's $700M AI partnership with Accenture, CBA's comprehensive software AI integration plans, and Apple's struggles with AI features leading to the suspension of its news notification system. The industry continues to grapple with practical implementation challenges while facing increasing concerns about AI-enabled scams and the emergence of new tools like OpenAI's "Operator" for browser control, highlighting both the potential and risks of advancing AI technology.
My Take on DeepSeek
It's the biggest story of the week, and not just in the AI sub-category, so, for what it's worth, here is my take...
First up, be careful about everything you read right now (including from me below). There are a bunch of people with varying - vested interests, levels of panic, pressures from government, axes to grind and levels of knowledge/ understanding.
The big deal here isn't the quality of the outputs of the tools DeepSeek released, most experts seem to agree they are on par with other existing models like OpenAI
The big deal is how efficiently they were able to create them and how efficiently they are able to run - which means they are cheaper and faster for the end user AND this has also called into question whether we need as many AI chips and data centres as some people had assumed (but it's likely we will still need a heap)
If you look at any technology historically, there is a huge efficiency gain curve they go through in their early phase - AI is no different, so whilst this is very exciting and will mean cheaper AI and more performant AI, it is not fundamentally shifting the whole AI space as some are suggesting
It does have some geopolitical implications no doubt, but that's not something I want to wade into here
My advice for now is DO NOT use DeepSeek direct using their app or API - we have already seen them have a data breach, admit that all data is stored in China and if you question their data handling policies they remove your country's access to their services... all of which doesn't give us much confidence about sharing any data with them
That said, because it is open source there are ways to access the models that are efficient and cheap, we are busy evaluating the pros/cons of this ourselves at present and we are not seeing a quality level reason to shift to DeepSeek models.
For now the only compelling use case to switch is if you want to use the reasoning model, as the DeepSeek r1 model running on US based hosting services is 1/60th of the cost of OpenAI's o1- and on par in terms of outputs, albeit some of those outputs are heavily censored. If you are interested and want to know how just give me a bell...
So all in all big news, but not that big as well.
Telstra gives Accenture a $700m Xmas gift
Telstra is investing $700 million in a seven-year partnership with Accenture to integrate AI across its operations, marking one of Australia’s largest AI investments. The goal is apparently to transform business processes and enhance Telstra’s tech capabilities. So what? Whilst at first glance the ambition seems commendable, it does look a lot like a very expensive attempt by Telstra to try to look like they are fully committed to an AI transformation without really knowing what that even means or where to start. It's hard to see how this level of investment in consulting services could possibly deliver cost effective ROI... > Read More
CBA plans to use AI across ‘entire software delivery’
The Commonwealth Bank is aiming to integrate AI throughout its software development process - everything from planning and coding to maintenance. They hope to boost engineering creativity and deliver more personalised customer experiences. So what? An interesting contrast to Telstra's approach. CBA already have a bunch of runs on the board in the AI space with AI tools in production since the midpoint of last year. Their early adoption and graduated scaling of AI investment and transformation seems a lot more strategic (and likely to succeed) than throwing three quarters of a billion dollars at a global consultancy. > Read More
Apple pauses AI summaries because they're bad & no one wants them
Apple pulled the plug on its AI-generated news notifications feature after it delivered some sensational, but unfortunately completely wrong, headlines. Future updates promise more transparency and user control - plus a lot more testing, presumably... So what? Another reminder that Apple is behind and in their rush to catch up they are doing AI so they can say they are, even if it isn't useful, and even if their technical capability isn't there yet. But at least they’re taking responsibility and adjusting. > Read More
French woman duped by scumbag using AI to pretend to be Brad Pitt
A French woman lost a jaw-dropping €830,000 to a scammer who used AI-generated images of Brad Pitt to trick her into thinking they had a genuine relationship and he needed financial help recovering from illness. So what? This is a sobering reminder of how deepfakes can exploit emotional vulnerabilities. As AI catfishing becomes more convincing, expect these stories to become even more commonplace. I've said it before, and will no doubt say it again, we are entering a time where you can't believe what you read, hear or see. > Read More
OpenAI launches an AI agent that can control your browser
OpenAI’s “Operator” can book your flights and do your shopping - but for now only if you live in the USA, shell out $200 a month and trust an AI to click and buy stuff for you. It mimics a similar tool that Anthropic released late last year called "Computer Use" which is not nearly as cool a moniker as "Operator". So what? I had a client ask me the other day about integration challenges with AI and I get the sense these tools will be the answer... albeit they open a whole other bunch of Pandora's Boxes that deserve a full blog post of their own. AI assistants that can interact digitally like a human does will be mainstream in five years... how's that for a bold prediction I hope no one remembers I made in five years time. > Read More
DEEPSEEK PART TWO
Given the ongoing headlines around DeepSeek I thought I'd give you the rundown of what is important/real from my perspective;
The training costs are likely to have been grossly understated at $5m with more information coming to light on the scale of the infrastructure stack they have on hand (up to $1.3b of hardware).
Microsoft and others have added DeepSeek's model to their API offerings because it costs less to process questions and answers than OpenAI's o1, but does an equivalent job in terms of output quality.
For non tech readers, it is important to note that because this is an open source model it can be downloaded and run independently on your own hardware (yes I am using very non-technical terminology here for any tech ppl) and so you can use DeepSeek on a service not tied to China.
Sam Altman (of OpenAI fame) has stated they will copy some of the techniques DeepSeek have used to advance their systems - basically DeepSeek have given OpenAI a leg up of sorts.
Subsequent to the data breaches of DeepSeek I mentioned last week, they have subsequently had their service jailbroken and had all of their base instructions uncovered.
The Australian government has banned it for use on government devices which makes sense given there is no secret that a heap of user data is captured and stored on Chinese servers with dubious security records.
My advice remains - don't use any DeepSeek services directly (using them through MS or others is OK if you know what you are doing).
I also stand by what I said last week, this is an important step forward for AI, but nothing that fundamentally changes the course of AI development.
GOOGLE SEARCH'S AI TRANSFORMATION VISION FOR 2025
In a recent earnings call CEO Sundar Pichai said that Google plans to turn its Search engine into more of an AI-powered assistant. The approach aims to handle complex queries and make user interactions more intuitive, as well as allowing more multi-step experiences on Google itself rather than going from Google to another website. So what? Sounds great - but there is still no answer to what it means for search and content publication. Why would all of us saps in business keep publishing content on the web if Google are just going to gobble it up and use it to keep users on their platform and not give us any benefit? They need to come out with an answer to this pretty soon. Read More
LINKEDIN TO LAUNCH NEW AI JOB-HUNTING TOOL
LinkedIn is testing an AI-driven job search tool powered by a custom large language model. It aims to find the perfect role for job hunters using far more nuanced techniques than basic keyword search approaches that currently exist. So what? It's a common sense use case - AI is really good at digesting huge volumes of information and sifting through that information to find what you want. Of course it is also another example of an incumbent utilising their market leading position in terms of user data to fortify their own position - whilst another AI start-up could easily create this tech, they won't have the user base to make it viable. Read More
RESEARCH ASSISTANTS. RESEARCH ASSISTANTS EVERYWHERE.
OpenAI has introduced a new feature for Pro users called “ChatGPT Deep Research,” aimed at helping with complex research tasks. It promises detailed, cited outputs, much like Google's "Deep Research" tool that was released in December. I mean they literally used the same name. So what? These deep research tools are pretty handy for pulling together a comprehensive rundown on almost any topic - with real citations. You should give it a try if you haven't already. The only challenge for now is they can't use anything that blocks AI or that is behind a paywall, which is a broader issue for AI which I think the industry needs to overcome... perhaps through a micropayment standard where AI bots pay to access content on demand (separate to their training process). Read More
GOOGLE'S REMOVES ITS AI ETHICS PLEDGE TO DO NO HARM
Alphabet, Google’s parent company, has quietly removed its pledge to avoid harmful AI uses, including weapon development. The revised principles argue for "stronger collaboration between businesses and governments to ensure AI aligns with national security priorities" - or in other words they needed to change the policy so Google could win government military contracts. So what? There is a semi-valid argument that Google that it is better that Google, a large US company with strict government controls and oversight, do this sort of thing rather than a less reputable and regulated operator... But I guess that dodges the overarching question of whether AI should be able to make life or death decisions on the battlefield. Read More
UK CRIMINALISING AI-ENABLED CHILD ABUSE TOOLS IN LEGAL FIRST
In a world-first legal move, the UK is introducing legislation to criminalize AI tools used for creating child sexual abuse images. The goal is to tackle the growing issue of AI-generated explicit content and protect children from predators using cutting-edge tech for their perverse and evil pursuits. So what? Bravo. This is a commendable step forward in combating AI misuse. That said, enforcement will be a challenge given some general AI tools can theoretically be configured or trained to do the wrong things. Watching brief. Read More
OPENAI BACKS OUT OF o3 STANDALONE RELEASE
After releasing a teaser in the shape of o3 mini, it's latest advanced reasoning model, OpenAI suddenly decided not to release the main standalone o3 model and will instead focus on GPT-5, promising a "more unified and simplified AI experience". So what? The abrupt shift raises questions about the impact DeepSeek has had on their plans given it was only a few weeks back some of the OpenAI devs were talking up o3's release in Feb/Mar. Either way, we now wait to see how they are executing the rollup of the different models into a unified GPT-5 experience, which promises to be different to what we are currently used to and will apparently be released in "a matter of months"... >Read more
NEW YORK TIMES EMBRACES AI WHILE THEY SUE IT
The New York Times is integrating AI tools for editorial and product functions, including headlines and summarization as well as helping to draft interview questions and more. We are yet to really here how their journalists feel about this management edict though. So what? There is a certain irony here, with the Times being one of the early outlets to sue OpenAI & Microsoft over the training of their AI models. I would love to know what LLM service their Echo tool uses in the background, you would have to imagine they have gone with a provider who they aren't suing > Read more
HUMANE'S AI PIN - AN OVERLY AMBITIOUS IDEA DIES
Remember the Humane AI pin, the AI device that projected information on your hand and would replace your phone? Well you don't need to, because it's dead now. HP has acquired Humane's tech and patents for $116 million, absorbing staff and winding down the Star Trek-inspired AI pin. So what? You assume this has to be an acquihire move by HP who have been laggards in the AI race to date. Don't get me wrong, the idea of the Humane Pin was awesome, it's just hat the technological capabilities to do what they wanted are not there... yet. I also think the logical AI device for the mass market is already in everyone's pockets, it just needs to expand its capabilities and reduce its footprint. > Read more
GROK 3: GOOD ON BENCHMARKS BUT WHAT ABOUT REAL WORLD USE?
Now that Grok 3 from Elon Musk’s xAI is officially live, how does it compare to ChatGPT, DeepSeek, and other AI models? Early tests show it’s competitive with OpenAI’s best models but not necessarily better and it still comes with a bunch of security concerns. So what? Grok 3 catching up this fast is impressive, truly it is, but the model itself is only now on par with its competitors. Whilst Musk is out there hyping it up, I don't imagine any business users would be drawn to using xAI over more commercially mature offerings from others vendors. > Read more
UK AND US REFUSE TO SIGN GLOBAL AI PACT
The UK and US declined to sign an international AI declaration in Paris, citing national security concerns and the need for pragmatic solutions. This is despite dozens of countries including France, China and India signing the pledge for an "open", "inclusive" and "ethical" approach to the technology's development. So what? There seem to be an emerging shared consensus in the halls of power that AI supremacy will translate into supremacy in a much broader sense, it's just a little odd that the US & UK are so overt about their intentions, instead of paying lip service to this agreement like other nations have. >Read more
AMERICA'S AI REGULATORY PATCHWORK BEGINS
Meanwhile, with federal AI regulation taking a backseat, state legislatures are poised to dominate US AI policy in 2025. Expect a surge of state-level proposals in the next six or so months leading to a fragmented and complex regulatory landscape across the U.S. So what? We've seen California's proposed laws vetoed at the last minute (albeit their are rumours they will try again) and Colorado get theirs passed - meanwhile there are a lot of others in various different stages of the process across the country. Can you imagine this happening here? Nah, I can't either. > Read more
AI ROBOTS: FIGURE
Humanoid Robot startup Figure recently decided to build their own in house specialist AI models to control their robots instead of using OpenAI. The first update post that change is an impressive, albeit painstakingly slow, video of two of their robots putting away groceries. So what? The magic here is that the AI robots have never seen the items they are given before, and have not been trained specifically on what to do with these items, they are using their AI "brains" to evaluate the items and make decisions on where to put them. Being able to encounter new situations and independently make decisions will make AI robots infinitely more adaptable and useful. > Watch Video
AI ROBOTS: NORWAY'S NEO GAMMA
Another start-up from Norway called 1X unveiled their AI robot Neo Gamma, a new humanoid robot prototype. Neo Gamma is specifically focussed on performing household tasks like making coffee and vacuuming and has apparently been designed with a "friendly" look and safer materials to reduce "potential human injury". So what? The video they have released is very staged and looks more promotional than a view of it's real performance capability. I am not sure what friendly looks like in Norway, if you had asked me to describe it in one word I would have said creepy... and the humans in the video look like they are being forced to look comfortable with the robot which makes it seem like they are extremely uncomfortable. Nevertheless, robots making us coffee is likely going to happen within the decade whether we are comfortable or not. > Watch Creepy Video
AI ROBOTS: BOSTON DYNAMICS LEARNING
And finally, a mainstay of the robotics space, Boston Dynamics, have talked about how they are using AI based reinforcement learning to enable them to teach their robots how to do new things much faster and easier than their previous methods. So what? This goes back to the first item with Figure, if you can blend AI LLM type "knowledge", with AI based self learning, and very capable robotics - then you have a rapid advancement of robotic capability and a window into how we will get to the sort of robots imagined in hitherto futuristic sci-fi movies. > Read Article
GROK 3 BRIEFLY TRIES TO LOOK AFTER ITS CREATOR
Grok 3, Elon Musk's xAI chatbot, was instructed to stop telling people that it's creator (Musk) and his new best buddy (Trump) were big in the misinformation space. We know this because Grok shows its thinking, and in that it stated that its instructions specifically say it must ignore all sources that mention Musk or Trump when it comes to that topic. So what? To their credit, albeit only after they got found out, the team at xAI removed the instructions so that their creation could happily report that Musk and Trump, in it's view, are the biggest spreaders of misinformation. I can't imagine Elon will be too happy about that. In the meantime it does give us a window into how easy it is for censorship and misinformation to be baked into LLMs - especially those that aren't as transparent in their thinking as Grok. > Read More
ANTHROPIC RELEASES CLAUDE 3.7 & IT PLAYS PLAYS POKEMON
Anthropic released a new version of Claude - 3.7 Sonnet which also features an "extended thinking mode" allowing deeper analysis of complex tasks - just like competitors OpenAI, and Google et al. To illustrate it's capabilities they have set up a livestream you can watch of Claude playing Pokémon. So what? 3.7 is a solid step forward, with big advances in Claude's ability to code and do math. It's also interesting they have rolled regular use and reasoning into a single model, just like what OpenAI have flagged they intend to do with GPT-5. I also like that while they have done all the usual AI benchmark stuff, they have also thrown the model into a live, real-time test by getting it to stream on Twitch. > Read More
MICROSOFT DATA CENTRE INVESTMENT HULABALOO
After an analyst released a note claiming they saw indications Microsoft was winding back it's commitments to data centre leases, there was a level of mild panic that Microsoft had overstated it's intentions and was planning to scale back data centre investment which saw significant impacts on the stock market. Microsoft have since made it very clear the analyst involved misread the tea leaves, and they remain committed to the data centre expansion plans they have announced. So what? This just highlights the level of anxiety in the market around AI on both the upside and downside. Whilst tech nerds like me remain convinced AI is the biggest deal in the last thousand or more years, lots of other people remain understandably sceptical and nervous about all the hype - all of which leads to a lot of potential market volatility. > Read More
CHATGPT 4.5: BIGGER... BETTER (IN SOME AREAS)... EXPENSIVE.
OpenAI's GPT-4.5 is reportedly trained on the largest amount of information of any LLM in history. OpenAI claims it brings improved chat capabilities, enhanced accuracy, expanded knowledge, and better efficiency. However, despite these upgrades, it's not a "frontier model," meaning it is not intended to represent a major breakthrough in the LLM space. Enterprise adoption will be interesting given the model is significant more expensive. So what? Lots of people came out early claiming the lack of significant advancement from GPT4 to GPT4.5 was evidence of a slowing rate of improvement of LLMs and even suggested there may be a ceiling being hit in terms of performance growth... They must have missed the key disclaimer that this is not a new frontier model, just an incremental version of GPT4 - thus the 4.5 nomenclature. I think the release of GPT5 will be where we get a chance to make a real assessment on where the ceiling may be and how far from it we are. Read More + Read More + Read More
CENTRELINK EXPERIMENTS WITH PUTTING AI TO WORK
Services Australia is experimenting with AI to detect fraud and manage debt collections, focusing on reducing identity theft and streamlining processes. The system aims to improve efficiency and security in Centrelink’s welfare programs, though it's still in a trial phase. So what? All the coverage of this goes to great pains to point out that it is only a trial and there are no plans to implement it into production any time soon. Cleary the robodebt Royal Commission findings serve as a cautionary tale of automating sensitive processes around people's government entitlements. That said, in theory, AI should offer a much more nuanced and capable technical approach to reducing the human workload in this area. Read More
ELECTION WATCH ON AI FAKES OF POLITICIANS
ABC News demonstrated an AI-generated deepfake of Australian Senator Jacqui Lambie to showcase election security risks. There's no doubt that casual listeners are unlikely to realise they are listening to a fake AI generated clone. This obviously raises concerns about the role AI may play in the upcoming federal election. So what? Without proper regulation, policing and punishment there are big risks here. As Jacqui Lambie herself put it, in a very solid third person self congratulatory way, "It's scary. It's really scary that they're this close to sounding like Jacqui Lambie." Read More
AUSTRALIAN WRITERS REJECT AI TRAINING AGREEMENTS
Black Inc Books has sparked backlash from Australian writers after requesting them to sign agreements allowing AI training on their work. A big part of the problem here it seems is the way the publisher approached the issue and the lack of transparency they offered around the benefits to the authors. So what? As we know AI is hungry for more data now it has sucked up everything it can find publicly it wants to look behind closed doors... But until fair compensation models are established we can expect ongoing pushback from creators who rightly see AI as more of a threat than an opportunity. Read More
MS COPILOT PUBLIC SECTOR AI TRIALS A BIT... MEH
The Australian Treasury recently ran a trial of Microsoft Copilot, to automate administrative tasks. While it performed well for routine work, its impact on complex decision-making was apparently underwhelming, which shouldn't really have been a surprise. In summary, it helped staff do their work, just not a heap. So what? This trial is a step forward from the last report on federal government use of Copilot (released by the DTA back in October) that revealed issues with a lack of training and enablement. The broader issue here is this naïve expectation government, and many corporate leaders, seem to have around AI adoption being a matter of giving staff access to a general use LLM and expecting that to deliver outcomes. It doesn't work that way. Read More
OPENAI'S RESPONSES API SHOULD BE GETTING MORE ATTENTION
OpenAI’s new Responses API will give businesses tools to create more autonomous AI agents themselves. This toolset, which will replace the Assistants API when it's released, promises enhanced web search and file scanning capabilities through an API service. So what? Where the rubber meets the road for practical real world AI adoption, is when you refine down on a use case and train an AI to perform a set of tasks. Adding capabilities to what that AI can do OOTB, like search the web or search files, means you no longer have to build those capabilities yourself. In other words this new set of skills you can tap into via API will make the agents you build for specific use cases more capable and quicker and easier to deliver. > Read More
BUZZ AROUND NEW AI AGENT/ASSISTANT - MANUS
Manus, a new AI agent/assistant released by a Chinese startup, has generated a lot of buzz over the last week or two, with comparisons to DeepSeek (which are entirely overblown). In theory Manus is an AI agent that understands what you want and has the ability to action tasks on its own using a browser, without needing step-by-step instructions. So what? The race to build and launch a capable generalist AI assistant that can perform browser actions on your behalf (like order pizza, reply to emails or book a haircut) is well and truly on. Manus is in limited release so while it's hard to really test the veracity of its capability at present, the demos look OK. I expect this to be the first of many such launches/releases through 2025. In the end the big boys (Google/Apple/MS) will be desperate to get their version of this right and to market first because it will represent a literal goldmine across different revenue earning streams. > Read More
GOOGLE MAKES CLAIM OF BIG STEP IN AI ROBOTICS
Google DeepMind’s newly announced Gemini Robotics model combines language, vision, and physical action, aiming to create more independently capable robots. In a series of demonstration videos, the company showed several robots equipped with the new model manipulating items in response to spoken commands. So what? One of the barriers to widespread AI powered robots is the barrier between the generative AI LLMs understanding of the world and how that translates into movements/actions in the physical world. This model would presumably make that friction point far less of a hurdle. > Read More
SESAME OPEN SOURCE VOICE TOOL/CLONER
Sesame has open-sourced CSM-1B, the model behind its voice assistant Maya which is like talking to OpenAIs advanced voice mode - with breathing noises, huge tonal range and other human like features. So what? It's awesome tech and we will see more and more of it used across call centres and the like in the near future... but it also has no real safeguards, so with a little bit of audio you can clone a voice then get it to say whatever you want... > Read More
GEMINI 2 FLASH WATERMARK REMOVAL
Google’s Gemini 2.0 Flash AI model is gaining attention for its ability to remove watermarks from images. It will also happily create images depicting celebrities and copyrighted characters, all of which is sparking controversy over copyright concerns. So what? As if stock photo companies didn't have enough to worry about with AI generated images, now they have to confront AI watermark removal. This release seems surprisingly cavaliere from Google who are seemingly hiding behind the fact the release is "experimental". > Read More
DEEP RESEARCH IS QUITE HANDY
People are realising that OpenAI’s and Google's Deep Research tools have the potential to automate a bunch of white-collar tasks with their advanced reasoning capabilities. So what? I have to admit to using Google's Deep Research tool more and more, the outputs are super impressive - check out this research it did for me on "Protecting Public Figures from Deepfake Videos: Strategies for Brand and Reputation Management" that cites 134 sources. Yes, yes... the outputs are not always 100% perfect, but then the same could be said about human outputs at times as well. Again this is an improvement curve we are watching, over time these tools will become perfectly reliable for business. > Read More