code51

Weird thing is, Gemini didn't crack speech-to-speech yet. They have a product but both Anthropic and Google are visibly suffering in speech-to-speech vs OpenAI voice mode.

Actually Sourcegraph's "Amp Code" is testing out a free ad-supported coding agent. Here is a video showing how it works: https://ampcode.com/news/amp-free

"Supported by ads from developer tool partners we’ve carefully chosen"

It's not trying to secretly insert tools into LLM output but directly present the product offering inside the agent area.

At one point, I speculate that Cursor will test this out as well, probably in a more covert way so that tool use paths get modified. Once the industry realizes tool-use-ads, then we're toast.

I'm surprised these pockets of job security still exist.

Know this: someone is coming after this already.

One day someone from management will hear about a cost-saving story at a dinner table, the words GPT, Cursor, Antigravity, reasoning, AGI will cause a buzzing in her ear. Waking up with tinnitus the next morning, they'll instantly schedule a 1:1 to discuss "the degree of AI use and automation"

Exactly and this is hell for programming.

You don't know whose style the LLM would pick for that particular prompt and project. You might end up with Carmack or maybe that buggy, test-failing piece of junk project on Github.

Vibing, not firing at his ML problems.

He's doing a capability check in this video (for the general audience, which is good of course), not attacking a hard problem in ML domain.

Despite this tweet: https://x.com/karpathy/status/1964020416139448359 , I've never seen him citing an LLM helped him out in ML work.

I doubt he's letting LLM creep in to his decision-making in 2025, aside from fun side projects (vibes). We don't ever come across Karpathy going to an LLM or expressing that an LLM helped in any of his Youtube videos about building LLMs.

He's just test driving LLMs, nothing more.

Nobody's asking this core question in podcasts. "How much and how exactly are you using LLMs in your daily flow?"

I'm guessing it's like actors not wanting to watch their own movies.

Junior jobs will come back when blitz-pricing of AI coding products end. Current bosses think these prices with 200/mo to "leave it and auto-code for the whole month, day and night" will stay like this. Of course it won't.

Typical startup play but in massive scale. Junior jobs might come back but not in bulk, still selective, very slowly.

Anthropic is actually a good point to focus on since Claude is very good proof that it's not about the scaling. We are not quite there yet but we are "programming" through how we shape and filter the input data for training it seems. With time, we'll understand the methods to better represent.

Current situation doesn't sound too good for "scaling hypothesis" itself.

The underlying assumption is that language and symbols are enough to represent phenomena. Maybe we are falling for this one in our own heads as well.

Understanding may not be a static symbolic representation. Contexts of the world infinite and continuously redefined. We believed we could represent all contexts tied to information, but that's a tough call.

Yes, we can approximate. No, we can't completely say we can represent every essential context at all times.

Some things might not be representable at all by their very chaotic nature.

High probability your v2 voice will break with this.

Sometimes the play is only about the announcement itself, not the implementation - news about a possible big news - just like upcoming Trump tariffs.

How convenient that insiders jumped on INTC with this "rumor" 3 days ago.

They made about 25% from this in 3 days. Easiest money from a rumor.

Despite all its worth, Suno doesn't approach the problem well.

After a few laughs and cheap replicas, people realize that it's damn hard to produce a good sounding, creative piece. Suno almost always adds noise and you can feel most of this is coming from fingerprinting.

With Suno and Udio, you lose control. You can generate starters and helpers but sooner or later, you want real control. That's not editing a section and having a conditioned piece of garbage seemingly fitting to the rest. No, control means completely changing progressions, sudden but calculated change of beats, removing any instrument for the shortest time and putting it back with razor-sharp studio detail.

I know a few of these are already addressable, you can take the output, separate into channels (if it's simple enough), quantize, edit and have a good one. Yet, you're not really supported anymore. What should have been was these other core music production software to get cheaper and/or far more effective.

Suno and Udio is a top-down approach. Maybe one day Logic Pro, Ableton, Melodyne etc. will fill in the details up to this point, coming from the ground up with AI, I don't know. We're not there yet and it just brings down the mask of mainstream music industry with its all-repeating shallow beats marketed to hell. Hearing mainstream was awful but it suddenly got even more awful.

Much of this "bad actor" activity is actually customer needs left hanging - for either the customer to automate herself or other companies to fill the gap to create value that's not envisioned by the original company.

I'm guessing investors actually like a healthy dose of open access and a healthy dose of defence. We see them (YC, as an example) betting on multiple teams addressing the same problem. The difference is their execution, the angle they attack.

If, say, the financial company you work for is capable in both product and technical aspect, I assume it leaves no gap. It's the main place to access the service and all the side benefits.

With current state of legal, a real challenge can happen only around 10 years from now. By then AI players will gather immense power over the law.

Why is Apache DataFusion not there as an alternative?

Initially LLM researchers were saying training on code samples made the "reasoning" better. Now, if "language to world model" thesis is working, shouldn't chess actually be the smallest case for it?

I can't understand why no research group is going hard at this.

So it'll turn to yet another arms race - similar to captcha, cybersecurity and nuclear weapons. SEO will use AI to fill in fluff inside AI-generated content (which is already done).

It won't directly match ChatGPT logs and OpenAI would just be pouring precious compute to a bottomless pit trying to partial-match.

Are the previous investments counting as "donations" still? Elon must have something to say...

Landing page yes but is it actually working with production performance? I have strong doubts after executing your own demo query of "Summarize the reviews you have on products in the OCR category." in the system.

If the idea is just demo and collecting leads, then you could have cached at least your demo queries.

Everybody's too landing-page focused these days.

For Mac and iOS, you can install ChatGPT app.

Why they won't enable search for their main web user crowd is beyond me.

Perhaps they are just afraid of scale. With all their might, it's still possible that they can't estimate the scale and complexity of queries they might receive.

So it very well could be that thinking on a worn neuron, trying to do memory access on it, might damage it even more (leg analogy)? It could be possible that the "swelling" could be redirecting brain to use other paths but with caffeine you're utilizing a worn-out neuron.

Maybe with chronic deprivation + caffeine habit, you're weakening your cognitive function - repeatedly using a misfiring neuron. At least, with no caffeine, you'd feel your memory is hazy and hard to recall, go to sleep.

Both will seem dull going forward. TTS will feel more natural every passing year so the current spending for any TTS model will seem kind of wasteful after 1-2 years.

"in an all-stock deal"

https://www.reuters.com/markets/deals/boeing-agrees-deal-buy...

The economy expects us all to be guild navigators - doesn't matter how your body is transformed if you're producing some economic value in the near term.

Simple, looking at the mirror and saying "Google-killer" firmly 3 times everyday.

I don't know where Pieter pulls the data from but one should note that this doesn't cover all accidents.

Pegasus Airlines has an excellent standing in this list but there are accidents where the plane got off the track.

One of these (February 5th, 2020) actually ended up fatal with 3 people dead out of 183: https://tr.wikipedia.org/wiki/Pegasus_Hava_Yollar%C4%B1%27n%...

Yet fortunately in another one (January 13rd, 2018) nobody got hurt: https://tr.wikipedia.org/wiki/Pegasus_Hava_Yollar%C4%B1%27n%...

In both cases, the plane was a Boeing 737-800.

I'm assuming this page is counting crashes, not the accidents that happen on the track.

Here's a wild guess: Through SpaceX design experience (rip apart, challenge all assumptions and rebuild better), Elon Musk (Falcon Aviation) will try to build a far more efficient and safe alternative to existing 737s.

Falcon Aviation got hold of various aircraft, which I think they'll just reverse-engineer and throw out what's not working. It's like taking pretrained weights from an established LLM: https://www.businessinsider.com/elon-musks-spacex-bought-a-b...

With Trump coming to power, Boeing's tarnished credibility will be a major issue to address. I'm expecting US government will incentivize this process. Boeing will lose good engineers out to this fresh competition.

Any case similar to this? What kind of boards you find to be more fitting for the case of OpenAI? It may not be Tesla but you’re probably drawing another reference from history, right?

All women, announced on International Women's Day.

The only significant side that's obvious.