shaism

According to Boris Cherny from Anthropic [1], the threat of prompt injection has been largely solved.

[1]: https://x.com/bcherny/status/2086520950259118464

Awesome. Perfect for editing with agents and sharing with coworkers.

- Even works well on mobile.

- Embedding images and videos also works flawlessly.

What is your use case?

Have you considered moving to open source / Chinese models?

If gemini-2.5-flash-lite is good enough for your application, you will find even lower cost options with better performance outside of the Google ecosystem.

I worked in a startup whose value proposition was largely derived from using graph data and graph databases under the hood. The main benefit is, as the repo even states "Fast data processing: Links traversal is processed with O(1) complexity."

So, technically, you can do deep traversals quicker. A few notes:

1. Few use cases truly need low-latency deep traversal on realtime data (>5 hops). There are some well known ones like fraud detection in payment processing and, possibly, social media recommendation engines. But I am not even sure how latest social media engines work, and whether they still rely on graph DBs.

2. However, in practice the advantage is often marginal. With modern analytical databases, or even an optimized PostgreSQL (ltree, materialized views, pgrouting, pg_duckdb, etc), you often get more than good enough performance. In addition those traditional SQL DBs scale with hardware more easily than graph databases. So, you can always use the lever: "Throw more hardware at it."

3. Even Graph DBs don't get good traversal performance under all conditions without hand tuning. For example, there is the "super node" issue, a node with an abnormally high number of connections (edges). And if you still need hand-tuning, you might as well choose something more versatile.

4. The ecosystem of a PostgreSQL and other popular DBs is just unbeaten. With graph DBs, you often prematurely put yourself in a corner that you don't want to be in.

Hence, my recommendation. Unless you are really sure that a graph DB is the right fit for your use case, start with something else, and go the graph db way when you have established a true need.

I created a version for personal use a month ago.

https://github.com/SharifIsmail/tendril

Feel free to give it a try. Agents can also directly interact with it.

It is unfortunate for the employees but it is necessary, and hopefully also a wake up call for other German companies and the government.

It’s time for leaner structures, smaller teams, faster decisions, and shorter development cycles. Volkswagen, like no other company, symbolizes the lethargy of the German economy and business environment.

Hopefully, the seven years with barely any economic growth will end soon.

> A small task that cost 121,000 tokens done directly cost 513,000 tokens when fanned out to two subagents, because every subagent has its own bootstrap cost, and the parent then consumes its transcript.

Is that true? My understanding is that the subagent only returns the result of the request without the main agent consuming its entire transcript.

Which provider do you use to access the MI300x?

Thanks! There is something magical about those symbolic approaches. But in practice, they seem to often fall short of expectations. Wonder whether there will be another "symbolic AI summer" one day.

Thanks for sharing!

I see. Cool. So did you generate the inference / reasoning rules to map those 500k events to 40 by hand or are there solutions to automatically discover those rules based on past incidences?

What would be a good use case for this?

"Halle is in the state of Saxony-Anhalt, which has an election in September. The far-right anti-immigrant Alternative für Deutschland (AfD), which is leading the polls with about 42%, has seized on the row."

That is the important part.

Yeah. Agree. The workload changes towards more orchestrating, editing, and reviewing rather than writing.

I see the same for slide decks. The bar has changed. Expectations are higher, but the high quality decks are still heavily edited, reviewed, and created by humans.

Based on your message, I assume you are and want to stay in software engineering.

Having worked in regulated and very sensitive industries (banking and semiconductor), I can tell you adoption is slower but AI is still everywhere. Because those industries have many constraints, they have to think hard how to adopt it. So, you will encounter AI initiatives everywhere.

If you want to be part of those adoption initiatives but just want more thoughtful and slower adoption, such industries might be a good fit. If you want to avoid AI initiatives in general, I am not sure whether software engineering is the right function.

Ilya mentioned in the video that 2012 and 2020 was the “Age of Research”, followed by the “Age of Scaling” from 2020 to 2025. Now, we are about to reenter the “Age of Research”.

How do I get stuff from my “onesmartphone” to the “onecomputer”?

Or shall I also put the “onesmartphone” in the cupboard?

Do you mind elaborating?

I agree that weather data should be public but I don’t see why we should restrict innovation in the private market if there is demand for it.

Also more generally, I see no issue in the government outsourcing work to a competitive private market wherever possible.

Very cool. I implemented something similar for personal use before.

At that time, LLMs weren't as proficient in coding as they are today. Nowadays, the decorator approach might even go further and not just wrap LLM calls but also write Python code based on the description in the Docstring.

This would incentivize writing unambiguous DocStrings, and guarantee (if the LLMs don't hallucinate) consistency between code and documentation.

It would bring us closer to the world that Jensen Huang described, i.e., natural language becoming a programming language.

The answer almost certainly is no. While lithography is one of the largest single contributor to manufacturing costs, the contribution to overall cost is still far below 10%.

And one cannot simply substitute an optical lithography with a nano imprint machine without redesigning some part of the process (etch, metrology etc.).

Investing R&D resources for a (best case) 10% reduction in costs while still having a decent probability of failure in a big but declining node is not worth it.

Maybe the market hasn’t recognized the value yet.

Hence, buy $GOOG.

Given your requirements, I would rank the frameworks like this:

* SEO -> 1. Astro, 2. Remix, 3. NextJS

* Performant Pages (for SEO) -> 1. Astro, 2. Remix, 3. NextJS

* Easy to Maintain (for a small team) -> 1. Astro, 2. Remix, 3. NextJS

* Self Hostable -> 1. Astro, 2. Remix, 3. NextJS

* Some interactivity and smooth transitions -> 1. Remix and NextJS, 3. Astro

From your description, it seems you are most familiar with NextJS. My question: How about the other maintainers? If everyone is familiar with NextJS, that is a big bonus point.

In conclusion, it comes down to how high you prioritize "Some interactivity and smooth transitions" and "Familiarity with the chosen framework." Astro, being a multi-page framework first and foremost, will not deliver the smooth transitions that the other two will. However, it shines on pretty much all your other requirements.

If you can live with Astro being a Multi-Page Framework that, by default, does a full-page reload for every route, and your team is comfortable using Astro, I would go with Astro. Otherwise, pick NextJS or Remix, whichever your team is more familiar with.

Actually that is how it works in many companies in Germany. It has its benefits and drawbacks, like every system ;)

Because for most users the costs of getting used to a new OS outweigh the benefit of switching.

TSMC today has more than 60% of the foundry market share, and an estimated 80%+ market share for the leading edge.

If you exclude Intel themselves (although they also use TSMC now) and Samsung, TSMC is pretty much the sole supplier to the leading edge.

So, yes, TSMC has the capacity.

Your Boeing / Airbus analogy hinges on multiple factors. First, Airbus and Boeing have comparable capacity. Second, they have comparable products. Both are not true when you compare Intel and TSMC.

NVIDIA is heavily supply constrained. Why haven’t they sourced Intel or Samsung as second source?

Historically, there are also many other issues with Intel operating as foundry. Do you think NVIDIA and AMD will be happy to send their CPU and GPU designs to Intel for manufacturing? Independence was one of the main drivers why TSMC was founded. To have an independent supplier who does not compete with its customers.

Devils Advocate: If there is no disruption in Taiwan, and TSMC continues to execute as it has, and Intel continues to execute as is has (not), then Intel might go to $0 or survive only because of government subsidies and military contracts.

NVIDIA, AMD, Apple and other companies seeking cutting edge performance will choose the most advanced node to stay competitive. Being the second best foundry has historically not been good business.

This is an extreme scenario, of course.

I am working in search (not Google).

My take: Today, using AI for search-related problems is still not cost-effective for most use cases. That being said, the landscape is evolving quickly. First, in some areas, an individual search creates more value than in others. An individual consumer doing a Google Search is totally different from a lawyer searching for reference material. Areas where the individual search creates more value can already benefit from AI today. Second, LLMs become exponentially cheaper, driven by more cost-effective computing but also more cost-effective models. Look at the pricing of GPT4o-mini vs GPT4 (the original). The models are comparable in performance for many search-related problems, but the price has decreased by 200x in 1.5 years ($0.15 vs $30 per 1M token). If that price trend continues, more and more search use cases will benefit from AI.

I am not sure. While Intel still has a dominant market share in personal compute, it’s data center market share has been shrinking for years and Intel doesn’t have a convincing story why that would change.

The main issue I see for Intel is that they are trying to catch up in multiple areas at the same time against more specialized companies with more resources, and fend of new architectures like ARM.

Look at the segments:

- Foundry: TSMC is way ahead, has the right culture and is better capitalized. Even Samsung is ahead.

- CPU Architecture: Compete with AMD on the x86 side and with Apple, Google and the likes with ARM architectures.

- GPU: Compete against NVIDIA and AMD.

- AI/NPU: Compete with NVIDIA, AMD, Qualcomm,…

All those companies were smaller than Intel in the past, and Intel couldn’t keep their advantage. Now, those companies are ahead and in a better overall position.

Nothing is impossible but I can certainly see why investors are skeptical, especially given Intel‘s track record in the past few years.

Sure, there are limits. Not everything is predictable, not even physics. But that is also not the point of such a model. The goal is to forecast across a broad range of use cases that do have underlying laws. Similar to LLM, they could also be fine-tuned.

Fundamentally, the pre-trained model would need to learn a "world model" to predict well in distinct domains. This should be possible not regarding compute requirements and the exact architecture.

After all, the physical world (down to the subatomic level) is governed by physical laws. Ilya Sutskever from OpenAI stated that next-token prediction might be enough to learn a world model (see [1]). That would imply that a model learns a "world model" indirectly, which is even more unrealistic than learning the world model directly through pre-training on time-series data.

[1] https://www.youtube.com/watch?v=YEUclZdj_Sc