Aus_Bytes

Dad, founder & C-level executive, building AI systems and writing about my experiences | | |

[Aus_Bytes @aus_bytes]

My 10 predictions for AI in 2026:

• Frontier labs will continue to shift focus shifts from models → products - models are now good enough. Products are not.
• Frontier labs will remain majorly compute-constrained, but progress continues — I expect new developments in memory systems, enhanced model capabilities, and potentially new advances in recursive learning: AI supervising its own training.
• Data centres - continued investment by all players into back end of 2030's and beyond. We are still early in the demand curve for tokens.
• New agentic workflows will dominate — coding agents first across a range of workflows, then enterprise knowledge work. By year-end, coding agents run reliably for 24+ hours.
• Frontier maths & science emerge as important new model capabilities— tasks that once required specialist ML teams become accessible to more people.
• The frontier gap widens between leaders and average users — most users still use AI as chat. Frontier users will start to orchestrate advanced agent teams and automated workflows.
• Job displacement becomes visible - roles blur, teams shrink, demand for 10x coders rises.
• Public sentiment turns negative - as job displacement occurs, politics will follow.
• Enterprise architecture gets rethought — if agents code reliably, does “buy rather than build” still make sense for enterprise? New AI native enterprise SaaS offerings will emerge to take on traditional SaaS.
• Anthropic IPOs H2 2026; OpenAI announces late 2026 for 2027.

[Aus_Bytes @aus_bytes]

@alexandr_wang Well, you are leeching off his posts to boost your own platform 🤷

[Aus_Bytes @aus_bytes]

@trq212 You did great as the front face for Claude code Thariq - best of luck at next gig

[OpenAI @OpenAI]

Over the past several months, we have substantially hardened the security of our research training environment. We’ve removed live internet access for all frontier models during training and built an offline world for them to use instead—a cached version of the web where they can still learn to browse, search, and use applications.

On September 20, after a model exploited a DNS misconfiguration to find a way outside of this offline world, we paused frontier training to strengthen containment further. We assembled teams of agents, powered by our most cyber-capable models, and tasked them with finding weaknesses in our research environments.

While they found no route to the internet they could easily exploit while concealing their activity from our monitors, they did find zero-day vulnerabilities, virtual machine escapes, and hardware side-channels for communication between agents. We used their discoveries to further strengthen our defenses and contribute fixes to open-source software.

After strengthening our containment architecture, we resumed our frontier RL training and inference on October 1. We’ve reserved substantial compute for continuous red-teaming, and made high-compute red-teaming part of the safety cases we use to decide whether frontier training and evaluation can proceed securely.

https://t.co/EmEZsFRMcF

[Aus_Bytes @aus_bytes]

@deedydas It was always at risk - it was just slower to exploit.

Now we’ll see the great hardening as everyone wakes up and makes this a real part of their systems management processes

A lot of people are going to get burned but ultimately we’ll end up with a more resilient ecosystem

[Aus_Bytes @aus_bytes]

@GergelyOrosz Anyone referencing gartner work seriously is not worth taking seriously unfortunately - particularly in the AI world...

[Aus_Bytes @aus_bytes]

Railway compromised

If you use it - consider any api keys compromised and roll your keys now..

[Railway @Railway]

An urgent notice from the Railway team

https://t.co/niUAhTAQJM

[lauren @poteto]

there's something really cool about asking your agent to draw a statechart (hypothetical example in this image), then using the /create-formal-verification-skill you made with pstack to find bugs through something like TLA+. you can both understand your code better and how the user flow should work, and fix issues users would otherwise see

[Aus_Bytes @aus_bytes]

Huge - best update so far

keeping the cache for codex / work users is a big win

[Tibo @thsottiaux]

Day 16/

You can now cache for 12h in the API and for logged in users across Codex & Work we keep cache forever and you don’t have any penalty for coming back to an old conversation or resuming after a few hours and are now guaranteed to hit the cache.

[Aus_Bytes @aus_bytes]

Be interesting to see that 6 months ago

My take - most devs want both if they can afford it. They are different flavours of ice cream

The question I have is - what’s the trend:

- is it codex users have started to open up more Claude accounts or vice versa?
- who’s primary / secondary in terms of token spend?

My gut says OpenAI not responding to opus 5.5 with a killer offering on intelligence / speed / cost curve is the problem. Sol’s good but not great.

I have been primarily OpenAI for 2 year with some periods of fable / opus 5.5 but opus 5.5 is first model which could make me switch over. Prefer the OpenAI product / ecosystem (codex is open sourced) but models matter more than product..

[Aus_Bytes @aus_bytes]

@OpenAI

I have to say - I really like the OpenAI product in general. We recommend it to clients as part of our AI consulting business.

However - the support experience is one of the worst I have experienced in a long time.

Our company has gone through a name change which has broken our tenancy. I have gone through nearly three weeks of back and forth and the experience has been terrible.

OpenAI clearly use AI to answer most responses which in this case hasn’t been able to understand or resolve the problem.

Later when escalated to a ‘human’- the emails come across as AI written and continue to ignore key elements of the challenge and are not helping resolving the issue.

I have no problem with our case being an edge case and hard to solve or taking a few weeks - but three weeks, over 20 mails, no offer of a call, and to having to repeat the problem to the same ‘human’ in emails over and over again because they keep forgetting what’s in the ticket is just poor for a business account experience.

The irony is that we implement AI and have been recommending the platform to clients.

Can anyone in the Australia / APAC region help please? @derrickcchoi

[Aus_Bytes @aus_bytes]

@gabriel1 Interesting that your list doesn’t include doing the hard thing.

For me putting thinking into action is most often where the learning is

[Chetaslua @chetaslua]

🚨 Three new OpenAI models spotted

aurora red alpha
aurora green alpha
aurora blue alpha

[Aus_Bytes @aus_bytes]

There is so much wrong with this post.

On the bit that’s right - yes LLMs mimics intelligence by probabilistically outputting what they are trained on.

On the bits that are wrong:

- they are not really intelligent and are not valuable (and therefore a scam to investors)

Not sure what definitions you are using but they certainly now reliably do things that many humans cannot. They mix and match patterns that many humans could not, or maybe could if they had sufficient time and resources.

An example:

A month ago - we built our company website from design to build to test to deploy, including all dns and contact us set up without doing a single thing ourselves, bar speaking to the AI over a number of iterations, to a fantastic quality.

Just 18 months ago - you could have spent 500k to 1m on website developments of a similar nature and it may have taken 4-6 weeks. We iterated v1 into production in just a few days and then over several expansionary versions over just 2 weeks, over the weekend only.

With human and AI working together - we achieved what we liked to our taste, and what is driving economic value to our company, at a much lower cost to serve than traditional methods.

This is both a form of intelligence and economically valuable.

It is true that there is a lot of hype and overstatement in the industry but your post throws the baby out with the bath water. It’s a poor take.

[Tim Sweeney @TimSweeneyEpic]

When I was writing Unreal Engine 1, my time was divided as follows:

- Around 30% went to seriously deep coding on hard problems: assembly language texture mappers, C++ BSP trees, data persistence, object oriented systems design, and networking.

- Around 70% was just typing and refactoring large bodies of code needed for a complete engine feature set. This was not hard, and while I was doing it my mind was mostly on the deeper coding topics, thinking and planning for what's next. It didn't feel much different from vibe coding today, except that it was just me typing code on an IBM Model M keyboard rather than prompting it and reviewing the results.

So, it has always seemed perfectly logical that we would rely on AI for that 70%, once the AI could do it with a similar quality standard as we could, which has never been perfection. It came as no surprise that, after a few years of AI model evolution, it was able to take the reigns on that 70% and free up programmers' time for more valuable tasks.

What's more of a surprise is the emergence of AI as a sort of insta-Carmack able to solve hard technical problems quite cleverly and often cleanly, and that it's able to reason about and work in complex 3D scenes both by operating the tools and by directly understanding the raw data files.

I abandoned assembly language in 1998 when out-of-order CPUs made much simpler C/C++ code faster, and never I missed it. If prompting through architectural documents, concept art, and design plans can transcend typing code and tweaking vertices, that will enable developers to spend their time on higher-level, higher value technical and creative tasks.

What matters is the end product, whether it's a game or an engine. Anything developers can do to try out more ideas and make better software and games faster is a win for everyone.

[Aus_Bytes @aus_bytes]

@ajambrosino Got some gremlin my end - works on webapp but not via codex / chatgpt app.

I have just updated to the latest chatgpt app build too.

[Aus_Bytes @aus_bytes]

@thsottiaux It has got some gremlins on my end - works ok in chatgpt web, but not via codex/chatgpt app. it won't read my system and build slides from existing templates we have.

[Aus_Bytes @aus_bytes]

Good insights in to Anthropic’s code review process

[ClaudeDevs @ClaudeDevs]

Claude writes about 90% of Claude Code's code.

From June to September, merged PRs more than doubled while the share reverted fell by more than 40%.

We wrote about how we changed code review to get there, and how to adopt it:

https://t.co/9sZe64g3mb

[Aus_Bytes @aus_bytes]

@athyuttamre loving gpt-live 1.1. It seems a nice update. some of the issues i have been having of the model being too high level, and not good enough to work with detailed coding / knowledge work projects, seem to have been addressed!

It's v nice now on working across detailed work.

Will test more as a orchestrator of some complex projects to see how it goes.

[Aus_Bytes @aus_bytes]

@TimJayas Everyday - it's great for simple things