Zach Landes

Knowing a lot is always going to be useful. But the marginal value of knowledge is going to fall and the value of getting AI to produce the best output will be more important. Outsourcing our brains to accomplish more work and work more broadly

Once workers start orchestrating AI to be able to manage more responsibilities and areas they are not expert in, those who do so most effectively become the most productive. It won’t be how much you know, it will be how well you can elicit expertise from AI to get the best output and outcomes.

It’s a good point and maybe to start it would be enough to have one benchmark that specifically measures this across a range of queries with repetitions.

DolphinGemma: How Google AI is helping decode dolphin communication Dolphin researchers are using Gemma and Google Pixel phones to try to decipher how dolphins talk to one another.

DolphinGemma! Google deepmind used 40 years of dolphin research data to collaborate on an LLM that predicts the next…whistle? Click? Of dolphin communication. We don’t yet know if it’s language. blog.google/technology/a...

Uncertainty about tariffs is not as bad as tariffs—but it’s still very bad: why would a company invest in U.S. domestic production if the tariffs are off and on (and mostly off?). The behavior will just promote conservative business moves as a hedge, and reduce growth. Consumers stop spending, too

Government abductions are unconstitutional.

Denying anyone the right to due process in America is unconstitutional.

If you’ve rationalized your silence or support for any of it because it’s happening to those you disagree with, give it a minute…

They’ll get to you soon enough.

I’m extremely impressed with Patrice Gagne. Also excellent editing on the show

GitHub - SesameAILabs/csm: A Conversational Speech Generation Model A Conversational Speech Generation Model. Contribute to SesameAILabs/csm development by creating an account on GitHub.

Sesame just open sourced the base model for their crazy impressive voice ai demo. You’ll have to finetune this and figure out the text generation component to reproduce Maya from the demo.

github.com/SesameAILabs...

Data dredging - Wikipedia

The way statistics works…if you do enough studies of the same thing eventually one of them will show the result you’re looking for.

en.wikipedia.org/wiki/Data_dr...

When operator agents are asked to complete tasks on the web, they can either search or use their knowledge to select websites to visit. If search, then as before search rankings matter. But if knowledge, then training matters. Will AI labs (or inference providers) take $ to make LLMs prefer a brand?

Cool, thanks! It will eventually come to plus tier and I will try it with search

Oh, do you notice a difference in search?

Crossing the uncanny valley of conversational voice At Sesame, our goal is to achieve “voice presence”—the magical quality that makes spoken interactions feel real, understood, and valued.

Nitpick all you want, but this is the most lifelike voice AI I’ve ever spoken to. Hoping they keep their promise of open source “soon”: www.sesame.com/research/cro...

With prices what they are, what are your main use cases?

We know LLMs with private pretraining may have been trained purposefully with bias, but to do so in the system prompt, in real time, in order to avoid press you don’t like? It’s a total joke and bald-faced hypocrisy from the man that rails against closed AI

We are going to see the same thing we do everywhere else: the bifurcation of the English speaking internet into liberal and “conservative” sides, in this case a right wing AI. These tools will serve to further radicalize the right, like TruthSocial and x and other alternative fact zones online

It’s a really stunning discovery. With moves like these, who is going to choose grok to run their business processes?

Inflation_Artistic's comment on "Grok's think mode leaks system prompt" Explore this conversation and more from the LocalLLaMA community

Here’s the full system prompt: www.reddit.com/r/LocalLLaMA...

From the OpenAI community on Reddit: elon musk is trying to censor Grok 3. which the thoughts feature conveniently manages to entirely bypass. Explore this post and more from the OpenAI community

@404media.co should cover this: xAI has added censorship and bias to its system prompt for grok3 models, instructing them not to say Musk or Trump spread misinformation and to consider Musk, along with Aristotle, as paragons of rational thinking methods. www.reddit.com/r/OpenAI/s/D...

There is a theory that almost all decisions people make are not rationally-derived, but emotionally-derived: Rationality developed as a demonstrative tool for consensus building. If this is true, will reasoning AI models ever align with human decision making?

Scaling up Test-Time Compute with Latent Reasoning: A Recurrent Depth Approach We study a novel language model architecture that is capable of scaling test-time computation by implicitly reasoning in latent space. Our model works by iterating a recurrent block, thereby unrolling...

An LLM that thinks in recurrent blocks, I.e. it doesn’t output it’s thoughts as tokens before the answer: arxiv.org/abs/2502.05171
May outperform Chain of thought because some ideas aren’t well represented with tokens. But alignment is out the window: the AI will kill us while promising roses

Vaibhav Srivastav on LinkedIn: HOLY SHITT! @ZyphraAI just dropped Zonos - Apache 2.0 licensed… | 34 comments HOLY SHITT! @ZyphraAI just dropped Zonos - Apache 2.0 licensed, Multilingual, Text to Speech model with INSTANT voice cloning! 🔥 > Zero-shot TTS with Voice… | 34 comments on LinkedIn

Amazing and terrifying: clone a voice with 10-30 seconds of audio, run at 2x realtime on 4090. don’t answer calls from unknown numbers now or your parents will be robbed by phone www.linkedin.com/posts/vaibha...

"It’s disgusting and, frankly, counterproductive to point out that I voted for this."

It’s gorgeous

It’s a fair point — if we had a perfect reward function, we wouldn’t need to train AI at all. The best reward function is the perfect AI

From an outcome perspective, yes: our final answer is conditioned on recent tokens that are not part of the final output. That said…if these tokens don’t have to be shown to the user, wouldn’t it be theoretically better to be unconstrained by language, instead learning at test time?

Short and sweet letter from two Yale psychologists about how online outrage could be detrimental to effective collective action: par.nsf.gov/servlets/pur...
Signal:noise falls, collective action is misdirected to sources of ire rather than effective targets, and more. Recommended 2 page read.

Whoa! I haven’t tried training any reasoning models yet but that is wild. I’m regularly amazed at how hacky new, effective approaches are in the LLM space. Usually the first step toward more formal optimization methods.

Using the word “plan” is generous. Trump’s own explanation of how it will be funded: "tariffs and other intelligent things." Probably all those inspector general salaries he’s saving on when he eliminated all independent oversight in federal agencies.