Sorry team. This is a step backwards. It’s woke as hell. The veneer of fluff its belching is unhelpful, and it’s determined to rewrite things chewing more tokens, and takes putting it in a headlock to get it to not drift for any structured outputs over 4k. You were on the right track with 5.0. This isn’t it.
America is in debt due to incentives. Want more votes, spend = give people money, project over budget add a tax. If we fired people for missing budgets or say shutting down government for weeks. Like we do in the commercial sector. I’m pretty sure we’d be able to stay out of debt for the most part.
@BarackObama Have you seen the actual current lines drawn by the democrats. It’s not a power grab, Republicans just own rulers and what the democrats put in place.
@radekosmulski@jeremyphoward Ok I’ll help you here. Ipynb is just a json file. Rename your notebook to .json prompt your changes and rename it back. For extra convenience write yourself a .sh tool call to handle the naming back and forth. You’re welcome.
Lots of words to say “china has a program to buy things for poor countries we should too”. Difference is China builds ports, USAID built drag shows. Panama was ultimately a flex to say, prob should think twice before taking that money. Stop over intellectualizing this. Not that hard.
I agree that most of these have downward pressure combined with austerity policies. The real question is what happens as it forces reciprocal tariff renegotiations with those who tariff the US. Medium term this may end of being closer to a balance of trade which should be a net positive impact.
@PalmerLuckey Last tip. Metadata creation is the way. Have an ontology layer, topic meta etc. Remember all data is time series, so add consideration for that too
Write a parser to turn them all into .md files, add meta data, link images etc the same way or just had a model describe the image. Add search function ideally structured not vector, attach to LLM and capture feedback. Approach 2. Place add md files into obsidian and create an obsidian MCP.
@PalmerLuckey Also anyone saying use long context doesn’t understand context windows, how these are being managed (long/short term memory, rolling windows, summarization or any of the other hacks to say context is unlimited) or training for most models. Usable context is at best 128k
Politics aside. Finally an AI pov that is connected to reality. This isn’t the first disruptive innovation, but it could have been the one that the collective hive mind crippled before it had a chance. Time for less hypotheticals and more hard work on real code. LFG
Deepseek is cool, highly unlikely their numbers are true, regardless two things happened last week.
1 - Towards the end of last year most folks had resigned to the fact that frontier innovation could only come from those with huge teams and unlimited compute. Today there's probably 1000 new teams fired up trying to do something new.
2 - This only pushes forward the OpenAI, Google, Meta, Anthropic towards faster releases. Over the last two months we've watched evidence of Claude throttling context windows, OpenAI making similar adjustments without changing pricing or really informing customers. While I get the need for them to get to profitability, this isn't cool. Models like Deepseek, Qwen and others keep the ecosystem honest.
Last note on the Deepseek topic. We would never place a foreign model anywhere near production, or dev for that matter. Even a 1% risk of of it invoking something is a non-starter for us. Foreign actors have continued to show impact with upstream attacks and IMO this is a pretty good way to do that. That being said the Chinese research papers are always good to read. Definitely add them to your reading list.