Ben Dickson

Software Engineer | Tech analyst | Thinker | Student of life | Founder of @bdtechtalks

[Ben Dickson @bendee983]

This is what I think the AI research community and industry need. In the past few years, we've seen a lot of progress, but most of the focus has been on transformer-based LLMs (and a few non-transformer-based variants) and next-token prediction.

But LLMs and NTP are not the only ways to achieve AI. I'm very excited about these new directions, including what @RichardSSutton is exploring and the world models that @ylecun proposes.

We need to run experiments in multiple directions. Diversity wins in scientific discovery.

[Richard Sutton @RichardSSutton]

I can’t say enough good things about John Carmack @ID_AA_Carmack and his Keen Technologies. But now Khurram Javed @kjaved_ and I have broken away to start our own startup and pursue a slightly different path toward understanding intelligence. Like Keen (and like Ineffable) we at Oak Lab @oaklab_ai believe in reinforcement learning and that intelligence is created and maintained from run-time experience. But we think current deep learning methods are weak and inefficient, and need not more tweaks, but fundamentally new ideas and a thorough reworking before they can provide a solid foundation for achieving the more ambitious goals of AI.

[Ben Dickson @bendee983]

@i_mika_el Part of being a good engineer is knowing which code needs your absolute attention and which parts you can let the AI do the work by itself.

[Ben Dickson @bendee983]

I think reviewing AI-generated code should not be a binary switch.

As software engineer, you should at least be fully aware of the interfaces, contracts, modules, and high-level architectural decisions that the AI coding agent is making. For more sensitive operations, you should absolutely review the code.

This doesn't mean that you should review every line of code. But you also shouldn't maximize speed by just providing a goal and letting the AI agent do all the underlying thinking and decision-making.

[antirez @antirez]

It is my belief that many devs right now are not maximizing what they can do with automatic programming because they still look at the code. Doing it makes you the bottleneck. Your time is better invested in new ideas, QA, design, and asking yourself what is your goal.

[Ben Dickson @bendee983]

@WellPaidGeek @DesignCntrl Exactly. I'm not talking about individual experiences. I'm talking about the broader market dynamics.

[Ben Dickson @bendee983]

AI coding agents will enable software engineers (and AI engineers) to do more with less

In existing markets, this means that companies can

1) either do the same amount of work with fewer engineers who use AI
2) hire more engineers (e.g., double their staff) and do 10x more with AI

But another effect of the falling cost of software is that many industries that previously couldn't afford bespoke software can now hire small engineering teams that can build software for them.

In other words, the market for software will expand with the help of AI.

Bottom line: Software engineers will be in more demand, not less. It might take some time to get there, but it will happen.

[Sam Altman @sama]

so far at least, i'm pretty sure AI has been net job-creating.

this was not what i expected--although i was much less pessimistic than others, i thought by this level of capability we'd have seen some impact.

it is possible this direction keeps going!

[Ben Dickson @bendee983]

@theslowtell Friction is an important part of learning and developing intelligence

[Ben Dickson @bendee983]

This is an interesting point. Failure -> recovery -> should be part of any advanced AI system.

But we can also look at it from a different perspective. If human intelligence relies on this loop of trial and error, what happens when we delegate all that work to AI?

There are a lot of discussions around AI doing the menial work and humans doing the high-level thinking and decision-making. But the problem is that to think, you have to "do." So without at least doing part of the menial work (e.g., coding, writing, etc.), what learning opporutnities are we losing?

[François Chollet @fchollet]

The process of trying, failing, updating a mental model, and trying again is the core of intelligence. We should celebrate models that fail gracefully and adapt instantly.

[Ben Dickson @bendee983]

The more AI models become advanced, the more they widen the gap between experts and non-experts.

Right now, experienced software engineers are starved for access to models like Fable and GPT-5.6. They can get very good results and productivity out of these models.

Engineers who are not caught up in the mentality that only closed AI models are useful can even take this a step further and determine how to use multiple closed and open LLMs to get the best results on the cost/accuracy frontier.

But for most average users, a $20 subscription to one of the three leading model providers is more than enough because they're not using the full power of AI (nor do they know what is possible with the frontier).

[Ben Dickson @bendee983]

@JedFrankowski2 Which is why demand will move to other sectors

[Ben Dickson @bendee983]

@innuendo_pibara 1.5x is about right (my experience too).

[Ben Dickson @bendee983]

One of the worst things you can do is let AI write your social media posts. The point is to express your thoughts in short form. If you can't exert that much mental effort, then something is wrong with you.

I use social media as my mental scratchpad. I usually just one-shot what comes to my mind, often don't even edit it and put it out there to get feedback from other people.

The process of formulating my thoughts, putting them out there, and getting feedback from people is a learning loop for me. When you replace that effort with AI, you are effectively denying yourself a learning opportunity.

[Polymarket @Polymarket]

NEW: Study reveals LinkedIn is the most AI-saturated major social platform, with over 40% of long-form posts flagged as fully AI-generated “slop.”

[Ben Dickson @bendee983]

The part where I agree with the most:

"Humanity has flourished through individual weirdness and creative tension. We envision alignment as a feature not of a single model but of an ecosystem of AIs raised in different places, disagreeing, competing, and learning from each other. We believe in keeping the weirdness alive."

When a single company (or a few very rich organizations) determines the direction and outcome of AI research, we lose the diversity and randomness that lead to innovation and discovery.

Progress requires participation by as many people and different opinions as possible.

[Mira Murati @miramurati]

Today we share the worldview behind our mission.

Human values don't average out. Local knowledge can't be centralized. The good future has many AIs, raised in different places, shaped by the people they serve, disagreeing with each other the way we do.

https://t.co/A14SurOM2K

[Ben Dickson @bendee983]

Every frontier lab thinks they should be the sole arbiter of artificial intelligence and no one else can be trusted with research and development.

The rest of the world thinks AI should be a collective endeavor, scientists should share findings and data, and models should be open source.

I'm not a frontier lab, so I definitely believe in the shared vision.

[Ben Dickson @bendee983]

@RoboIntellect You're going to need enough human-annotated data to be able to create an LLM-as-a-judge. It's very tricky and messy and requires bespoke annotated data.

[Ben Dickson @bendee983]

I've been following AI model routing. There are several solutions/frameworks. They all have two key characteristics:

- They work really well for verifiable tasks with a clear feedback signal (e.g., coding)
- They are extremely poor at handling unverifiable domains (e.g., creative writing)

Benchmarks are mostly verifiable tasks but real-world applications often revolve around tasks that need human feedback.

Will wait to see how this new LLM router works, but I don't think there is an out-of-the-box solution that will solve model routing for every application.

The best way for enterprises to create their own router is to have internal evals and annotated data that can help develop an LLM-as-a-judge model (whether through prompting or fine-tuning) for their specific application.

[Bindu Reddy @bindureddy]

🚨 Announcing Smart Model Router - Create Your Own Custom Router To Route To Your Favorite LLM

- Route to GPT Sol, Muse Spark, Grok 4.5 and Fable
- Optimize for cost or performance
- Use Fable only for hard coding

Once you create a router, you can use it on our Chat, Agent or in an API!

Coming on Monday - Support for all CLI clients including Code, Claude Code and others

[Ben Dickson @bendee983]

Hot take: FDEs are a trap. What the AI market needs is “decoupling engineers.”

On paper, AI companies like OpenAI employ forward-deployed engineers to help enterprises better adopt their models for their internal applications. In reality, FDEs lock you into the ecosystem of the company they work for.

Decoupling engineers are independent AI engineers that help your business use the best mix of models for your applications. This could be a mix of API-based frontier AI models and even self-hosted open LLMs. They make sure you’re not dependent on any one AI company.

Never trust an FDE who has a stake in making you use AI from only one provider.

You should not put all your eggs in one basket.

[Ben Dickson @bendee983]

The ironic twist in the current state of AI:

- U.S.-based Frontier AI labs use data created by everyone over the years to train LLMs that they release as very expensive closed models
- Chinese labs distill knowledge from those models to create their own competing LLMs and release them as free open source models
- U.S. Companies fine-tune Chinese models for their applications to avoid paying the premium of closed frontier AI

[Ben Dickson @bendee983]

@johncrickett Agreed

[Ben Dickson @bendee983]

Totally agree with John's take here. The framing should not be whether AI is good at X. It should be whether you're good enough to get AI to do X.

Most LLMs have a lot of internalized knowledge and can produce very good results in many fields/tasks. But to get the best of those capabilities out of the models, you need to:

1- Be an expert in that field (or at least know enough to be able to express your intent with enough clarity and detail)
2- Be good at prompting LLMs or building the right scaffolding/harness around them to get them to produce the right results.

Otherwise, you're just casting your line in a random location in an ocean of knowledge. You have to know where to fish.

[John Crickett @johncrickett]

Asking whether AI is better than the average engineer is the wrong question.

AI doesn't do anything on its own. It sits there until someone drives it.

The right question: who's driving?

Give AI to someone who can decompose problems, think in systems, and verify their own work. They'll increase their output.

Give it to someone who can't tell good from bad. They'll ship bugs faster.

The skills that separate these two groups aren't AI skills. They're the fundamentals: clear thinking, knowing what "done" looks like, building the right thing instead of just building things.

AI is a multiplier. And multipliers are brutal. They amplify whatever's already there.

If you're leading a team, stop asking "will AI replace my engineers?" Start asking "which of my engineers will AI accelerate?"

Then ask "how do I bring the rest of the team up to that level?"

[Ben Dickson @bendee983]

This will increasingly become a recurring pattern: You explore new applications and use cases with highly capable (and very expensive) closed frontier AI models.

But as your requirements and problem space become clearer and you reach production stage, you’ll look for more affordable (and sometimes faster) alternatives to closed AI.

While open LLMs might not match the full capabilities of Fable and GPT-5.6, they are more than enough for most use cases, and they are orders of magnitude cheaper (aside from all the other benefits such as self-hosting and customization).

[Sam Hogan 🇺🇸 @samhogan]

we're helping a customer spending $60k/mo move from OpenAI & Anthropic to open source models

they use almost every model offered by the labs, so we needed to find replacements for all of them

after generating evals, this is what we landed on

new cost: $12k/mo, 80% savings