K. Reaz

Cybersecurity Architect | PhD | Building Quantum-Ready Security Solutions

[K. Reaz @kr34z]

@moritzkremb waitlist

[K. Reaz @kr34z]

What a fantastic turnout of talents at the @TheOfficialACM #MobiSys 2024 conference in Tokyo. Thank you for your cordiality and kindness.

[BMFTR @bmftr_bund]

📄🖊Deutschland und die USA unterschrieben heute eine #gemeinsameErklärung, welche die gemeinsame Arbeit an #Quantentechnologien stärkt. Die #Zusammenarbeit soll ermöglichen, die #Schlüsseltechnologien möglichst schnell auf die Straße zu bringen.

[K. Reaz @kr34z]

@BVG_Kampagne BVG security there told me I can submit the letter and receipt online. I kept checking that https://t.co/yt0ZgVOJoQ is under maintenance. I am on my vacation, How can i avoid the fine? (3/3)

[K. Reaz @kr34z]

@BVG_Kampagne Went to S Bahn office to get replacement card and a letter confirming that I had a valid ticket when I was checked. Then went to BVG Alexandarplatz office and found the Kilometer long line, due to my connecting DB Train I could not wait. (2/3)

[K. Reaz @kr34z]

@BVG_Kampagne controller could not read (faulty card) @SBahnBerlin Firmen Abo Card, gave 60 Euro fine receipt. kept telling me I should go to BVG office, even though simple sense knows that I need to go to S Bahn office to replace card. (1/3)

[MIT CSAIL @MIT_CSAIL]

#otd in 1962 MIT released CTSS, the first functional time-sharing system that allowed multiple people to use a computer terminal at the same time.

[Andrew Ng @AndrewYNg]

Inexpensive token generation and agentic workflows for large language models (LLMs) open up intriguing new possibilities for training LLMs on synthetic data. Pretraining an LLM on its own directly generated responses to prompts doesn't help. But if an agentic workflow implemented with the LLM results in higher quality output than the LLM can generate directly, then training on that output becomes potentially useful.

Just as humans can learn from their own thinking, perhaps LLMs can, too. For example, imagine a math student who is learning to write mathematical proofs. By solving a few problems — even without external input — they can reflect on what does and doesn’t work and, through practice, learn how to more quickly generate good proofs.

Broadly, LLM training involves (i) pretraining (learning from unlabeled text data to predict the next word) followed by (ii) instruction fine-tuning (learning to follow instructions) and (iii) RLHF/DPO tuning to align the LLM’s output to human values. Step (i) requires many orders of magnitude more data than the other steps. For example, Llama 3 was pretrained on over 15 trillion tokens, and LLM developers are still hungry for more data. Where can we get more text to train on?

Many developers train smaller models directly on the output of larger models, so a smaller model learns to mimic a larger model’s behavior on a particular task. However, an LLM can’t learn much by training on data it generated directly, just like a supervised learning algorithm can’t learn from trying to predict labels it generated by itself. Indeed, training a model repeatedly on the output of an earlier version of itself can result in model collapse.

However, an LLM wrapped in an agentic workflow may produce higher-quality output than it can generate directly. In this case, the LLM’s higher-quality output might be useful as pretraining data for the LLM itself.
Efforts like these have precedents:
- When using reinforcement learning to play a game like chess, a model might learn a function that evaluates board positions. If we apply game tree search along with a low-accuracy evaluation function, the model can come up with more accurate evaluations. Then we can train that evaluation function to mimic these more accurate values.
- In the alignment step, Anthropic’s constitutional AI method uses RLAIF (RL from AI Feedback) to judge the quality of LLM outputs, substituting feedback generated by an AI model for human feedback.

A significant barrier to using LLMs prompted via agentic workflows to produce their own training data is the cost of generating tokens. Say we want to generate 1 trillion tokens to extend a pre-existing training dataset. Currently, at publicly announced prices, generating 1 trillion tokens using GPT-4-turbo ($30 per million output tokens), Claude 3 Opus ($75), Gemini 1.5 Pro ($21), and Llama-3-70B on Groq ($0.79) would cost, respectively, $30M, $75M, $21M and $790K. Of course, an agentic workflow that uses a design pattern like Reflection would require generating more than one token per token that we would use as training data. But budgets for training cutting-edge LLMs easily surpass $100M, so spending a few million dollars more for data to boost performance is quite feasible.

That’s why I believe agentic workflows will open up intriguing new opportunities for high-quality synthetic data generation.

[Original text: https://t.co/zOiuUFtmo3 ]

[K. Reaz @kr34z]

its an absolute honor to be part of this year #Harvard President's Innovation Challenge. Can't wait to find out the winner. #PIC2024

[K. Reaz @kr34z]

Who needs curated playlists to focus while studying when there’s ‘Background Noise’ in #iOS? So handy!🙏

[K. Reaz @kr34z]

Now Billions are being spent to teach machines so that we can learn from them later. #AI #ML #NLP

[K. Reaz @kr34z]

@BVG_Kampagne @BVG_Ubahn Was ist Ihre Ausrede, um kurze Züge auf der U3-Linie zu fahren, wenn die Nachfrage groß ist und die S-Bahnen nicht fahren? So verlieren Sie die Unterstützung der Publikum. lustige Plakate helfen nicht.#GDL

[K. Reaz @kr34z]

True respect to her excellency Raychelle Omamo for her empowering panel discussion and speaking from the bottom of her heart on behalf of the "Global South" #MSC2024

[K. Reaz @kr34z]

Full week at #CISPA, the epicenter of information security. Reconnecting colleagues from #USENIX. Space is the new playground of #Cybersecurity. Thank you @JamesPavur 🙏

[K. Reaz @kr34z]

An eye opening talk on Cybersecurity in Space. @CISPA summer school,thank you 🙏

[K. Reaz @kr34z]

@Apple : let me charge my AirPod wirelessly with my iPhone, please! Or tell me WHY you won't!

[K. Reaz @kr34z]

@ElonMuskAOC Because it has "Space Age" strength and style

[K. Reaz @kr34z]

#BVG: biite, keine kurzen Züge auf der Linie U3 zu den Hauptverkehrszeiten, insbesondere wenn die FU Berlin Semester fährt und es Sommer ist. Ich musste heute 3 Züge am U Heidelbergerplatz um 10:00 Uhr auslassen. Alle ankommenden Züge waren Kurzzüge und völlig überfüllt
#berlin

[K. Reaz @kr34z]

@BVG_Ubahn keine kurzen Züge auf der Linie U3 zu den Hauptverkehrszeiten, insbesondere wenn die FU Berlin Semester laüfr und es Sommer ist. Ich musste heute 3 Züge am U Heidelbergerplatz um 10:00 Uhr auslassen. Alle ankommenden Züge waren Kurzzüge und völlig überfüllt

[K. Reaz @kr34z]

@vodafoneservice bitte tun etwas. warum am Frankfurter Allee Berlin gibt KEIN funk! nicht nur heute sondern es ist so hier