bob1029

OpenAI Presence is available to eligible enterprise customers as a deployed product through a limited general availability program. Deployments are led by OpenAI Forward Deployed Engineers and select global systems integrators. Presence is not yet available as a self-serve product.

Building your own enterprise chatbot with in-house domain experts is the only path that makes sense to me. The software piece is really not that difficult. There are a lot of examples and options to pull from now. You could maybe implement a custom MCP server and use the M365 copilot on top if you don't want to reinvent the wheel.

I think bringing in OAI consultants is probably a mistake in most cases. I've already seen one AI consulting team and they just cannot get deep enough fast enough. It would take them years of suffering our codebase and daily procedures to get to the point where they could actually make an impact.

I have learned that the purpose of a CEO in many (most?) companies is not to actually run the company.

Their purpose is to be the human delegate that represents the corporate entity. They are empowered to act on behalf of the entire organization. This is important because a corporation of many individuals generally cannot act as one. You can sometimes get close with really, really good teamwork, but that's still not enough for a lot of customers.

If you are in some regulated B2B context, not having a CEO or having a "non-traditional" CEO can result in lost deals. The counterparties are looking for "one throat to choke" when they are being made responsible for approving large contracts. They want a person's cell phone number they can call when they're having a bad mood about whatever they've just paid for. They don't want to talk to your developers or the cranky old founders. They want to talk to a person who can be summoned to any arbitrary golf course within 48 hours and who will arrive in a fantastic mood 100% of the time. This person often doesn't have to know a goddamn thing about what your company's product actually does. They just have to be a delegate. A conduit of information and a representative element.

The primary design goal of the stealth mechanism is to get to the target undetected. By the time a radar signature shows up, the aircraft have ideally already dropped their bombs.

Being detected too late might as well be the same as not being detected at all in most practical scenarios.

To improve determinism, define a system instruction with explicit rules for your specific use case.

What if I want to do the other thing? When performing research with many sub agents, having a lot of diversity in the hypotheses is a big deal. If my 5 parallel sub agents all produce the same conclusion I might as well have only ran one.

The latest OAI models have done the same thing. I'm currently adding random variation to prompts to compensate for the lack of higher temperature sampling.

KUHF used to run classical music programming nearly exclusively up until 2011. Since then it's been relegated to their HD-2 channel and ran by some computer algorithm. If all you have is an old analog radio there is genuinely nothing to listen to around here.

Quite the straw man we've constructed here.

I am the exclusive user and I only watch from browsers on laptops and desktops attached to whatever display and audio outputs. I suspect this describes a non-trivial portion of the audience here as well. I don't care about someone being able to watch my media collection on Xbox because I don't have one.

I just did. Put the above into your favorite chat bot and append your preferred language, runtime environment and block storage provider.

There's never been a better time to consider writing your own media server solution from scratch.

I really don't like the idea of transcoding on demand. It has always created a janky experience on the client. I've been using a custom solution that runs everything through ffmpeg normalization as part of the media upload phase. Whenever a client wants to stream something we give them raw byte ranges directly from the block storage provider. In practice, this feels like flipping the channels on an analog tv. Every video starts and seeks nearly instantly.

I've long since given up on the dream of maintaining lossless master copies of everything just in case. Every practical use of that 50 gb copy of Home Alone would have you transcoding it. Most clients would choke on that bit rate. When would you actually be able to use all of the available information? Why not just discard what you'll never use? I realize it feels offensive to kill information like this but you can do a much better job with the end experience if you do.

You can get a modern chatbot to produce this experience in a single prompt. The code can run on anything slightly more powerful than a potato because you aren't transcoding at playback time anymore. A raspberry pi would be more than enough to serve a few clients on a LAN. If you remain humble in your list of demands (a simple directory browser that plays videos linked within), you can more easily achieve total autonomy.

I grow my own peppers (Serrano/habanero). It's very easy to get sufficient yield out of a few small plants.

The peppers at the grocery store are of grotesque proportion. The jalapeños look like bell peppers now.

Not enough humans to deal with physical machines and the compliance issues around them.

I agree if we are operating a cat meme distribution business that AWS starts to look a bit overkill. Something like a banking consultancy is a different matter. Good luck passing due diligence as a bank's vendor when you are also physically responsible for the machines. You can do this, but it sucks and it's your entire job + overtime + on call.

AWS has a dedicated feature (artifact) that can serve as the foundation of your compliance package. If you are racking your own machines you get to start with a blank pile of forms.

I've got overhead transmission lines in my back yard right now. It's really not a big deal. I've got a garden and a small shed that are directly underneath the lines. I have no perceptible loss in utility. Every now and then, someone will visit my property but I don't have to interact with them.

The nature of the easement can make a big difference. Oftentimes the power lines will be colocated with gas pipelines which tend to impose stricter requirements on land owners. It's a lot easier to do things like excavate when you can see what you aren't supposed to hit.

I have a hard time seeing this adopted in many markets due to overall system complexity and weird new edge cases.

Trailers are supposed to be pretty dumb. This is what makes them valuable from a logistical perspective. If we hyper focus on reducing the carbon footprint of each specific shipment we may find we created an overall increase in aggregate once the higher order and deferred consequences come to bear.

To be even more nuanced, the land is not typically taken entirely. The more typical solution is an easement with compensation for the land owner to cover the reduction in utility.

Transmission lines are pretty harmless when considering things that could be built instead. Zero emissions, silent, no illumination, etc.

Nothing feels better than mastering a specific workflow in Blender. After using visual studio all day, it's like a breath of fresh air to click through a scene and have the tool respond like it actually cares about UX.

I find it amusing how ineffective all of this stuff becomes if you simply leave your phone at home. No one is doing the hard kind of police work anymore.

The judgment of task difficulty is a cursed problem. In most domains you really don't know what is down there until you start digging a little bit.

Simply starting in the right part of the search space is the biggest predictor of success.

The best way to save tokens is to start out the deep research pass with cheap models and then funnel the findings through increasingly powerful models. I've got a pipeline right now that uses all 3 of the gpt-5.6 model variants to address each stage of the process. If you are using models like sol or terra to generate hypotheticals and perform initial exploration, you are leaving money on the table.

5 hypotheticals out of luna will massively outperform 1 hypothetical out of sol, but the cost is the same and so is the runtime if you do it in parallel. The hypothesis generation phase is also a great place to mix and match models from different vendors. The more diversity at this step the better.

The other thing I started looking into is batch pricing which represents 50% off for OAI tokens right now. With some tweaks to the UI/UX of an enterprise chatbot, I think it is possible we could have users get comfortable with the idea that questions to the robot might take a day to come back. The key is that this has to actually work. Users don't mind trading time if their questions come back with high quality results.

People have been automating WoW for a generation using things like peripherals duct taped to oscillating fans despite multi-million dollar budgets designed to defeat things far more sophisticated than this.

I would think of headless browser automation in exactly the same way you would about cheating in FPS video games. The red team always has the initiative and can win if they want to spend enough time and money.

If evolution had access to TSMC's process technology, do you think it wouldn't leverage it? The signals in my AMD CPU propagate ~1,000,000x faster than the ones in my brain.

See also:

https://arxiv.org/abs/2406.19108

We show that when random, non self-replicating programs are placed in an environment lacking any explicit fitness landscape, self-replicators tend to arise. We demonstrate how this occurs due to random interactions and self-modification, and can happen with and without background random mutations. We also show how increasingly complex dynamics continue to emerge following the rise of self-replicators.

The hypothesis generation phase is looking like the most critical part of having an agent reliably hit your targets.

Simply starting in the correct part of the search space is probably the biggest predictor of success. Forcing one big loop to fight its way through all the hypotheticals from zero looks like a dead end for many practical scenarios, regardless of how powerful the model is. I think you could draw some analogies to humans here.

I have found that delegating deep research to a simple tool call is the best way to ground the agent in complex domains. If you make the main agent loop carry the weight of this research, it's going to do a really shitty job because of how the RLHF tries to preserve context and get an answer to the user quickly. As a tool, you may find the agent invokes multiple rounds of research consecutively without realizing it has incurred billions of tokens of consumption. Many of the tokens are wasted when generating independent hypotheses and subsequently investigating them, but the point is that you sampled 10-100x search space before getting serious about mutating the environment. The tradeoff seems worth it in a lot of cases. Correctness >> Time >> Money.

This is a fantastic way to quantify just how bad a consumer experience this is. Flat pack furniture tends to be very awful compared to alternatives that are also often cheaper and without assembly requirements.

Everything that has come in box from a store like IKEA (or assembled in the back from a box and presented as non-flat-pack) has lasted me less than a decade. I've got a bedroom set that was built out of proper materials and it's almost a century old.

The thing that pushes consumers toward ikea is the consistency and convenience. Most things you can load and cart home by yourself same day. Moving around furniture built by the Amish is a serious logistical challenge by comparison. Maybe you could solo it with a hand truck and some experience, but it's genuinely dangerous to move some things without help. If you aren't moving frequently, the appeal of disposable furniture begins to fade quickly.

That's a situation your employer sets up for you!

Who says you need a boss or other human coworkers?

Start your own consultancy and take 100% of the pie. The term "employee" evaporates in this context, as does an entire universe of lazy, negative tropes about work ethic, etc.

Statistical power comes from having many samples. I agree that having just one sample doesn't take you very far.

Better to find joy in other parts of the process!

I recommend the joy of helping the customer get what they truly want. Spend more time on client calls and intercept additional hot potatoes now that we have extra free time to work with.

The technology is a blank canvas. These LLMs are like 600DPI color laser printers. The customer cannot tell the difference between hand painted and LLM generated at even a short distance. Once you get your head out of the worry that not hand painting everything is somehow an abdication of your professional capacity, you may find this whole thing goes a lot better. Try to think of it more like a heavy equipment operator climbing into the cab of a Caterpillar D9. You could do the job by hand, but the customer would probably get upset with how long it takes. Why not just finish the damn thing? Fire the infernal machine up for a few minutes and then move onto the next task.

You may find that by using the big scary machines that you get to engage with far more interesting technology problems than if you had insisted on hand coding. The only real difference is that you get to enjoy the problems from the comfort of an air conditioned cabin rather than a hot, muddy field.

With sufficient information you can derive a signal even in the presence of overwhelming noise. Assuming the noise is not perfectly correlated with the signal this is always possible.

Schemes like GPS, CDMA and DSSS are based upon this concept. GPS in particular is quite impressive in its ability to recover information that is received below the thermal noise floor.

There are some unusual but otherwise plausible ways to reject heat at incredible scale without involving a lot of mass.

https://en.wikipedia.org/wiki/Liquid_droplet_radiator

Space has many downsides, but it also comes with a few advantages.

I close my browsers with prejudice. If it mattered I would have put it into an issue or committed it to SCM before I wandered away from the tab. Also, history exists. If the reason we are keeping our tabs open is because we've disabled browsing history, then I have a simple suggestion to make.

Walking up to my workstation in the morning and seeing all the trash from yesterday's efforts causes me meaningful loss in motivation. I get down to a blank desktop at the end of each day now.

Certainly the idea is difficult to execute on. However, there isn't anything about it that fundamentally says it cannot happen. I think Starlink is a pretty good example of how this could go.

The framing of the detractors always comes off like we intend to build Costco-style structures in space with robots tending to literal racks of unmodified server equipment running web app and credit card processing workloads.

I think the most effective part of this architecture is having multiple initial conditions to sample from.

We've been looking at using gpt-5.6-luna to do hypothesis generation at scale. Running many copies of something approximately as powerful as gpt5.4 just to get a sense of what options exist before we put a stick into the mud.

Single agent loop does not work very reliably for deep research. Especially in domains with complex tool calling and environments. You can get it to perform sometimes (often enough for a demo to work), but the team will rarely adopt it because they want it to work ~100% of the time, not ~40%. The anchoring you get with initial findings makes it really hard to get unstuck without user intervention later on. When all findings occur as part of a deterministic research pipeline (tool), things tend to work better at the edges.

I've been considering a three stage pipeline that does hypothesis generation => investigation => synthesis using luna => terra => sol. This is the first LLM family where I feel like we can actually use the full range.