John Gorman

DaTa

[John Gorman @jp_gorman]

@hooeem This type of post is beyond ironic.... Literally having this post defeats the entire purpose of the LLM coding models. The same way having humans even reviewing this defeats the autoresearch purpose. Karpathy's code literally embeds the "meta bitter lesson" for all ;)

[John Gorman @jp_gorman]

Naughty, Naughty #Google #Gemini - #GDPR In my opinion - consent bundling and bad privacy design defaults don't seem the best way to gain users trust

[John Gorman @jp_gorman]

Mixed feelings about #AWS #Kiro for a new project. It manages specs well, but outlines how explicit you need to be as it goes off the rails and the implementation is very poor if you are not super explicit. It has a catostrophic forgetting issue despite steering and feedback.

[John Gorman @jp_gorman]

More blatant lies from OpenAi about unified models…..
“Right now:
•Text in this chat → GPT-4o.
•Voice → Also GPT-4o (real-time variant).
•GPT-5 → Only in certain environments or behind limited rollout flags, not in all consumer-facing modes.”

https://t.co/qxfVcWZvbP

[John Gorman @jp_gorman]

#Llama4 has been the best Public Relations win for Gemini 2.5 #Gemini #LLMs

[John Gorman @jp_gorman]

@NotAdverse @jeremyphoward Nope - they picked the smallest models for comparisons and left state of the art out of comparison….some of the responses I’ve seen online from llama4 are tragic! The minimal bench improvements for massive cost investment means these may be forgotten by …. Tuesday ;)

[John Gorman @jp_gorman]

OpenAI's new vision model when it works is amazing. I got it to take a bottle of exputex (image of a box on my kitchen counter) and a crafted prompt about a female skier and it generated the skier holding the box as if it was an advert. #OpenAI

[John Gorman @jp_gorman]

Turns out it rendered on their servers but did not show in the App or the website. It only shows the images when you refresh the rendered images.

[John Gorman @jp_gorman]

Wow - Yet again OpenAI is scaling like a 1980's cubscout website. 10 attempts to create a simple image on ChatGPT Plus on it's new image model and not a single one rendered #OpenAI #GenAI #ChatGPT

[John Gorman @jp_gorman]

This truely is embarrassing for #OpenAI on so many levels. Their supposed "orders of magnatude better" 03 models are loosing to their 01 models (even "preview") and DeepSeek-R1 and GPT-4o #DeepSeek

[John Gorman @jp_gorman]

...on the other hand #deepseek not only found the items I was looking for but found ones that had injected affiliate links that got me deals better than the store was offering directly. #Google #Gemini "thinking" can't even integrate with it's own search engine.

[John Gorman @jp_gorman]

One really has to wonder how Google will win at AI when in all the time I have used it's search engine to search for limitless items, it has never shown me one add relevant to what I do or what I want #google #AI #Gemini

[John Gorman @jp_gorman]

WOW deepseek API - A maintenance page! Retro 90's https://t.co/j7U7makmbs

#deepseek #maintenance #503

[John Gorman @jp_gorman]

@iam_danlewis @carrigmat I got the 32B running on my M1 when my RAM was at 85% and it still ran at about 4-5 TPS - MOE for the win! Once I kill chrome I'm sure that will speed up ;)

May need a new smaller hardware spec for the Unsloth versions at 80% reduction in size.

[John Gorman @jp_gorman]

@AntonyStringfe2 @fchollet Have to admit I find the same. The nuggets it finds in the preamble are impressive, the self reflection even more impressive.

[John Gorman @jp_gorman]

DeepSeekV3 is pretty fantastic for an open model. What's even more is the free usage on https://t.co/cA5qO0z9Di - I had to buy some API tokens given they are currently way cheaper than OpenAI GTP4-mini and it's ranked better too on LlmSys Leaderboard!

[John Gorman @jp_gorman]

Not sure if it's by the book, but Google is releasing features #OpenAI and #Grok and #Meta have been afraid to in the EU and allowing us to taste them for free. I'm loving their #AIStudio Top Job #Google nice to see you back!

[John Gorman @jp_gorman]

All the OpenAI talk about alignment means nothing if they can't release their models because they don't "align" with EU laws. Trading privacy for denial of service is not security #OPenAI #GDPR #DOS

[John Gorman @jp_gorman]

@sama Except the losers (sorry your friends) in the EU! Our gpt+ should be discounted for half the product features we loose out on because you can’t “align your models” to actual reality. What is the point in having “alignment” if it can’t align to the existing laws.

[John Gorman @jp_gorman]

@AdamRodmanMD Glad you say O1-Preview because the O1 experience for me under GPT Plus saw it get faster to respond, dumber and less reasoning with less medical knowledge in it's reasoning. Just my opinion.