deepGem

I built something similar at a hackathon, a dynamic teleprompter that adjusts the speed of tele-prompting based on speaker tonality and spoken wpm. I can see extending the same to an improv mode. This is a super cool idea.

Isn't that such a great outcome. No more robotic presentations. The best part is that you can now practice Improv at the comfort of your home.

A random malicious president ? Who was democractically voted by more than 70% of the country ?

Thank you. This is really nasty. Boxtown residents should sue xAI and take them to court.

Skepticism is valid. The environmentalists came after dams too.

I meant Europe is an example of how not to do regulation. The problem you just mentioned. If you get land easily electricity won't be available and vice versa.

This is a huge one. What Musk is looking for is freedom from land acquisition. Everything else is an engineering and physics problem that he will somehow solve. The land acquisition problem is out of his hands and he doesn't want to deal with politicians. He learned from building out the Memphis DC.

Am using Claude code as an approximation here. 2 years down the line the tooling around analytics will get integrated in AU assistants and they will be absolutely able to figure out unused features.

What will eventually pan out is that senior devs will be replaced with junior devs powered by AI assistants. Simply because of the reasons you stated. They will ask the dumb important questions and then after a while, will even solve for them.

Now that their minds are free from routine and boilerplate work, they will start asking more 'whys' which will be very good for the organization overall.

Take any product - nearly 50% of the features are unused and it's a genuine engineering waste to maintain those features. A junior dev spending 3 months on the code base with Claude code will figure out these hidden unwanted features, cull them or ask them to be culled.

It'll take a while to navigate the hierarchy but they'll figure it out. The old guard will have no option but to move up or move out.

They had 9.99 for the first year.

There is no competing product for GPT Voice. Hands down. I have tried Claude, Gemini - they don't even comes close.

But voice is not a huge traffic funnel. Text is. And the verdict is more or less unanimous at this time. Gemini 3.0 has outdone ChatGPT. I unsubscribed from GPT plus today. I was a happy camper until the last month when I started noticing deplorable bugs.

1. The conversation contexts are getting intertwined.Two months ago, I could ask multiple random queries in a conversation and I would get correct responses but the last couple of weeks, it's been a harrowing experience having to start a new chat window for almost any change in thread topic. 2. I had asked ChatGPT to once treat me as a co-founder and hash out some ideas. Now for every query - I get a 'cofounder type' response. Nothing inherently wrong but annoying as hell. I can live with the other end of the spectrum in which Claude doesn't remember most of the context.

Now that Gemini pro is out, yes the UI lacks polish, you can lose conversations, but the benefits of low latency search and a one year near free subscription is a clincher. I am out of ChatGPT for now, 5.2 or otherwise. I wish them well.

Product doesn't see the point of engineers being engaged and feed the engineering team like an in-house outsourcing shop.

Because they want to feel superior as the ‘this was my idea and you executed on my idea’ nonsense. Their answers to most ‘why are we doing this ?’ ‘trust me bro’. I am perhaps generalizing and there are outlier product managers who have earned the ‘trust me bro’ adage, but most haven’t.

This PM behaviour will never change. Engineers have said enough is enough and are now taking over product roles, in essence eliminating the communication gap.

Precisely.

I would any day take chatGPT/Claude over an IBM consultant. I worked at IBM.

"I don't "understand" how LLMs "understand" anything."

Why does the LLM need to understand anything. What today's chatbots have achieved is a software engineering feat. They have taken a stateless token generation machine that has compressed the entire internet's vocabulary to predict the next token and have 'hacked' a whole state management machinery around it. End result is a product that just feels like another human conversing with you and remembering your last birthday.

Engineering will surely get better and while purists can argue that a new research perspective is needed, the current growth trajectory of chatbots, agents and code generation tools will carry the torch forward for years to come.

If you ask me, this new AI winter will thaw in the atmosphere even before it settles on the ground.

This means there's a certain inertia: it can be better to handle the interim reports the same, even if they've been biased one way for several years, than to introduce a change that makes the numbers not comparable to history.

This is a very interesting point. So if BLS suddenly became more accurate, all the agencies have to re-tune their own biases and corrections => Could lead to short term discrepancies.

What one sees as inefficiency is actually efficient from a totally different lens.

Create punishment system? Unless compaies report data back to BLS very fast, they pay big fee or are taxed higher. Small shops would hate it.

Or incentivize companies to report accurate data pretty fast. Payroll management systems can be plugged in real time, but that costs money and yeah small businesses are not going to be happy. So incentivization works better than punishment I think.

The earlier reports are intentionally more noisy because there is value is being fast and then revising later, and everyone who uses this data is aware of that.

I get it and yeah my tone is very exaggerated. I don't think anyone in BLS should be fired and whoever is suggesting that does not understand how public institutions work.

I am just curious why there is so much of a discrepancy. This has been pretty much the status quo in BLS for a long time. They issue numbers and then they revise them later. However, you'd expect the revision to be moderately within an error %age.

Also how will this retroactive change help everyone involved. Ok, the new job numbers reflect a gloomier past (or a more vibrant past) how is that even helping everyone who is so focused on 'what's going to happen tomorrow'.

I retract my stance about BLS being intentionally corrupt - that's uncalled for.

What I really fail to understand - how can departments like BLS screw up to this extent. Either they are grossly incompetent or they are intentionally corrupt.

The data covers the period from March 2024 to March 2025 and trims the average monthly jobs gains seen during this period (roughly the last 10 months of Joe Biden's presidency and the first two months of Trump's) from a monthly average of 147,000 to about 71,000.

50% error. This is more or less consistent. How can a department have this error % and still have their job. I understand the data collection mechanism is not the most sophisticated, but even accounting for that, this consistent error % is not to be overlooked.

I wonder why there is such lack of accountability from firms whose data pretty much feeds the world's economy.

I spent a few weeks trying to build an alternative to self attention that scales memory linearly. I I got surprisingly good results. While in principle this makes a lot of sense, I am struggling to push the test accuracy above 86%.

Some of the alternatives I am about to consider:

1. Diffusion with sparse attention layers. 2. Hierarchical diffusion - next token diffusion combined with higher order chunk diffusion.

Still figuring out the code and I would love any feedback on these approaches.

I first thought this is about Kevin Kelly. Then somewhere midway I thought I was reading an autobiography. It was only towards the latter half that I realized this is the author talking about Kevin Kelly and visiting his house.

Even though the language is very simple, the writing is quite convoluted.

Currently,

You train a VLA (vision language action) model for a specific pair of robotic arms, for a specific task. The end actuator actions are embedded in the model (actions). So let's say you train a pair of arms to pick an apple. You cannot zero shot it to pick up a glass. What you see in demos is the result of lots of training and fine tuning (few shot) on specific object types and with specific robotic arms or bodies.

The language intermediary embedding brings some generalising skills to the table but it isn't much. The vision -> language -> action translation is, how do I put this, brittle at best.

What these guys are showing is a zero shot approach to new tasks in new environments with 80% accuracy. This is a big deal. Pi0 from Physical Intelligence is the best model to compare I think.

Ditto with my daughter. I don't really know if this helps them get into programming but she now types faster than me with much better accuracy.

I guess I'll stick to drinking water. But I'm sure there's a reason why that's bad for me.

Here's one :)

Too much water also erodes your body of salts. If you already are on a low salt diet this could be a problem. For us Indians eating mountain loads of salt, this is a non issue.

Doesn't this imply that the future of AGI lies not just in vision and text but in tactile feelings and actions as well ?

Essentially, engineering the complete human body and mind including the nervous system. Seems highly intractable for the next couple of decades at least.

Waymo clearly stands out as an exception amongst all moonshots that Google went after. However, they don't seem to have that one axis advantage in Waymo. I can't believe they didn't double, triple down on their efforts to build a fully integrated car. Compared to Apple, they were at a much better position to do this because of all the underlying tech/models and research.

May be that's the problem - that there is no one rallying individual for Waymo. They should just spin it off and make it an independent private company and retain % ownership.

I somehow feel Google will be way better if it's run like Berkshire, the CEO just focuses on capital allocation and let's the managers do their jobs in their respective companies - YT, Waymo, search, cloud, deepmind.

I'm not sure that culture can dissipate in Google at this juncture.

I've thought long and hard about free trade. Ultimately, people want cheap products and free trade optimizes for price that's it. It's like this single objective function that disregards all other functions.

So the competitive advantage for any country now boils down to the entire nation optimizing for cost. A country like China can do this, because the cost optimization for USA = lifting someone out of poverty in China. They are at the rock bottom of cost already.

Now that China has learned to optimize for cost, they will continue down that path - AI, robotics whatever tool they can amass to drive down cost, they will.

Meanwhile, other nations arbitrage about quality of life, politics, hatred of the rich so on and so forth and let the cost optimization slide. Ideally if every country optimized for cost, free trade would work like magic. Sadly, China is the only country doing that and rest of the world is now beholden to them, losing their once cherished competitive advantage.

Now China which once was derided for cheap quality has learned to build quality at scale and for cheap. Literally no one can compete with them on price:quality ratio. 3-4 decades of cost optimizations have paid off immensely for them.

I don't know why there is this general vibe amongst technologists that they are somehow far superior in their thought process compared to the general populace. I have heard the words peasant class come up quite often in the tech circles. Why is this ? Aren't you part of the general populace ?

At the risk of getting downvoted let me posit a different narrative.

What if the past governments were just propping up growth by increasing government spending. The facts that are coming out now point in this direction.

Unemployment numbers that were fudged and recalled by nearly 1M on two occasions. 35% government hiring in the last 3 years.

What if this is the unwinding that was needed long ago but is happening now.

Yes tariffs are a net negative for the US economy, but there is no guarantee that tariffs will stay as we saw with the Canada and Mexico tariffs.

So may be, the government is not completely incompetent.

This data from Apollo management points to a not so bleak picture.

https://www.apolloacademy.com/wp-content/uploads/2025/03/DC_...

This is also why devs are not in charge of purchase decisions at tech companies. I don't mean FAANG but the usual tech shops. Someone buys a tool and you have to use it. I think the startups selling dev tools are not selling to developers at all, but to the IT folks of these big tech firms.

Should they pull it off, it's not at all a bad startup to build. However, you need to now invest in a sales force that can sell to the Fortune 500. As a tech founder with no sales trope, this will be incredibly hard to pull off.

I digress, but yeah selling to devs is almost always a terrible idea since we all want to build our own stuff. That spirit may also be waning with the advent of Replit agent, Claude code and other tools.

Any idea what the sRAM to uRAM ratio is on these new GPUs ? If they have meaningfully higher sRAM than the Hopper GPUs, it could lead to meaningful speedups in large model training.

If they didn't increase the memory bandwidth, then 512GB will enable longer context lengths and that's about it right? No speedups

For any speedups You may need some new variant of FlashAttention3 or something along similar lines to be purpose built for Apple GPUs.