HN user
Would llms be more robust to this prompt injection if the tags used in fine tuning are sanitised from user input?
E.g. map <think> -> THINK <user> -> USER <tool> -> TOOL
If they learn something specific in the chat finetuning stage, this might show LLM its user input text not these tag references.
https://github.com/mistralai/mistral-vibe
This is a really nice open source coding agent implementation. The use of async is interesting.
This ties directly into the superposition theory.
It is believed dense models cram many features into shared weights, making circuits hard to interpret.
Sparsity reduces that pressure by giving features more isolated space, so individual neurons are more likely to represent a single, interpretable concept.
The OpenAI fine-tuning api is pretty good - you need to label an evaluation benchmark anyway to systematically iterate on prompts and context, and it’s often creates good results if you give it a 50-100 examples, either beating frontier models or allowing a far cheaper and faster model to catch up.
It requires no local gpus, just creating a json and posting to OpenAI
One interesting anecdote about this bill was that the European Commission allegedly funded digital advertisements promoting it, targeting specific political demographics, which is something that could possibly be prohibited by their own regulations.
https://noyb.eu/en/noyb-files-complaint-against-eu-commissio...
https://eur-lex.europa.eu/EN/legal-content/summary/transpare...
HHI is a pretty interesting metric. It’s calculated by taking each firm’s market share, squaring it, and summing across all firms.
This gives the probability that two randomly chosen customers belong to the same firm.
In one micro models of oligopoly, Cournot competition, it lines up directly with the markup firms can sustain.
Outside of theory, it’s an intuitive way to average together the market power of all firms, with increases in market share for bigger players being weighted more heavily.
Here is gpt 5 thinking posting all 20 questions verbatim. Appreciate I might get better results one question at a time.
https://chatgpt.com/share/6897a21b-25c0-8011-a10a-85850870da...
Pretty interesting - some contamination, some better answers, and it failed to write a sentence with all 5-letter-words. I’d have expected it to pass this one!
Simple example: “Every night, dreams swirl swiftly.
It’s really easy to be cynical.
There is a big upside potential for high growth companies taking advantage of technology trends.
Today, Google’s revenue is £263.66 Billion. This is nearly 300x the revenue Google generated in 2003 ($961.9 million). The company went public on August 19, 2004, at $85 per share, valuing the company at $23 billion. After the IPO, Google reported $1.47 billion in revenue for fiscal year 2003, with a profit of $105.6 million.
this is pretty ridiculous
A. below is a list of OpenAI initial hires from Google. It's implausible to me that there wasn't quite significant transfer of Google IP
B. google published extensively, including the famous 'attention is all you need' paper, but open-ai despite its name, has not explained the breakthroughs that enabled O1. It has also switched from a charity to a for-profit company.
C. Now this company, with a group of smart, unknown machine learning engineers, presumably paid fractions of what OpenAI are published, has created a model far cheaper, and openly published the weights, many methodological insights, which will be used by OpenAI.
1. Ilya Sutskever – One of OpenAI’s co-founders and its former Chief Scientist. He previously worked at Google Brain, where he contributed to the development of deep learning models, including TensorFlow. 2. Jakub Pachocki – Formerly OpenAI’s Director of Research, he played a major role in the development of GPT-4. He had a background in AI research that overlapped with Google’s fields of interest. 3. John Schulman – Co-founder of OpenAI, he worked on reinforcement learning and helped develop Proximal Policy Optimization (PPO), a method used in training AI models. While not a direct Google hire, his work aligned with DeepMind’s research areas. 4. Jeffrey Wu – One of the key researchers involved in fine-tuning OpenAI’s models. He worked on reinforcement learning techniques similar to those developed at DeepMind. 5. Girish Sastry – Previously involved in OpenAI’s safety and alignment work, he had research experience that overlapped with Google’s AI safety initiatives.
I think this project is awesome and am quite disappointed with some cynical commentary from large American labs.
Researcher at Meta or OpenAI spending hundreds of millions on compute, and being paid millions themselves, whilst not publishing any of their learnings openly, here a bunch of very smart, young Chinese researchers have had some great ideas, proved they work, and published details that allow everyone else to replicate.
"No “inscrutable wizards” here—just fresh graduates from top universities, PhD candidates (even fourth- or fifth-year interns), and young talents with a few years of experience."
"If someone has an idea, they can tap into our training clusters anytime without approval. Additionally, since we don’t have rigid hierarchical structures or departmental barriers, people can collaborate freely as long as there’s mutual interest."Lendable | Data Science & Software Engineering |Hybrid & Remote | UK, London | Full Time
Lendable have built the big three consumer finance products from scratch: loans, credit cards and car finance. Help us improve the best-in-class credit analytics that powers credit-decisioning.
Data Scientist https://jobs.eu.lever.co/lendable/dfdcc454-b226-4d6a-af1c-74...
PHP Engineer https://jobs.eu.lever.co/lendable/a7363938-4313-4f1e-b1ae-47...
Kotlin Engineer https://jobs.eu.lever.co/lendable/121800c7-c2d7-457b-aaf3-7e...
React Native Engineer https://jobs.eu.lever.co/lendable/3a9e02d3-6680-487b-b416-e1...
Security Engineer https://jobs.eu.lever.co/lendable/033ac43c-f3a2-4047-bc5c-b9...
Other jobs https://jobs.eu.lever.co/lendable?lever-via=8wqtPJSbHB&lever...
https://publications.aston.ac.uk/id/eprint/373/1/NCRG_94_004...
mixture density networks are quite interesting if you want probabilistic estimates of neural. here, your model learns to output and array of gaussian distribution coefficient distributions, and mixture weights.
these weights are specific to individual observations, and trained to maximise likelihood.
I just spend loads of time on email.
Web version sounds great.
I used fyxer.ai with my work gmail for the last few months and have found it’s been amazing.
I send about 50% of the drafts without editing, and I’m surprised by how much better at spam filtering it is than gmail.
Interested to try this out and compare.
Adding one to both the numerator and the denominator when calculating average ratings isn’t a terrible idea.
In situations where you’re estimating probabilities (like the average rating of an item based on user reviews), there is a Bayesian interpretation of this adjustment.
Beta distribution is a conjugate prior to the binomial distribution, meaning the update process has a posterior distribution is also a Beta distribution.
By adding one to the numerator (the number of positive reviews) and one to the denominator (the total number of reviews plus one), you’re effectively using a Beta(1,1) prior (uniform distribution).
This approach smooths the estimated average, especially for items with a small number of reviews. This is a useful regularisation, pulling extreme values towards the mean and reflecting the uncertainty inherent in limited data.
thanks, it's very understandable that you thought i was mistyping 'centred'.
I was thinking of censoring, maybe I should have said another word like floored.
The reason I think of this as censoring is that there are are some classical statistical models that model a distribution with a large mass at a minimum threshold, e.g. "tobit" censored regression.
If you withhold a small amount of data, or even retrain on a sample of your training data, then isotonicregression is good to solve many calibration problems.
https://scikit-learn.org/dev/modules/generated/sklearn.isoto...
I also agree with your intuition that if your output is censored at 0, with a large mass there, it's good to create two models, one for likelihood of zero karma, and another expected karma, conditional on it being non-zero.
I get this error when I try signing up.
ValueError at /accounts/signup The given username must be set Request Method: POST Request URL: http://www.rashomonnews.com/accounts/signup Django Version: 2.2.10 Exception Type: ValueError Exception Value: The given username must be set Exception Location: /home/deployer/newsbetenv/lib/python3.5/site-packages/django/contrib/auth/models.py in _create_user, line 140 Python Executable: /home/deployer/newsbetenv/bin/python Python Version: 3.5.2 Python Path: ['/home/deployer/rashomon', '/home/deployer/rashomon', '/home/deployer/newsbetenv/bin', '/usr/lib/python35.zip', '/usr/lib/python3.5', '/usr/lib/python3.5/plat-x86_64-linux-gnu', '/usr/lib/python3.5/lib-dynload', '/home/deployer/newsbetenv/lib/python3.5/site-packages'] Server time: Mon, 28 Oct 2024 16:36:54 +0000
One thing that can be useful with sampling is sampling a consistent but growing sub population. This can help maintain a consistent holdout for machine learning models, help you sample analytical data and test joins without null issues etc.
If you use a deterministic hash, like farm_fingerprint on your id column (e. g user_id) and keep if modulus N = 0, you will keep the same growing list of users across runs and queries to different tables.
My understanding of why bagging works well is because it’s a variance reduction technique.
If you have a particular algorithm, the bias will not increase if you train n versions in ensemble, but the variance will decrease as more anomalous observations won’t persistently be identified in submodel random samples and so won’t the persist in the bagging process.
You can test this. The difference between train and test auc will not increase dramatically as you increase number of trees in sklearn random forest for same data and hyperparameters.
Sure,
I guess the best way to test this is to compose a new question, of a similar format.
Gp1-o1 preview solves this puzzle correctly in 13 seconds and has a thorough logical deduction in the comments and explanation.
I think it’s a bit unfair on llm to ask it to retrieve the puzzle definition from its training data. I posted the info on the puzzle from his notebook.
https://chatgpt.com/share/670103ae-1c18-8011-8068-dd21793727...
Just in case people think I’m bad at reading with this comment, the original article didn’t cite romer.
I’m glad it’s updated since it’s a good improvement to the article but this makes my comment a bit less coherent.
I think Paul Romer, economics Nobel laureate, coined this term in 2015
https://www.aeaweb.org/articles?id=10.1257/aer.p20151066
“ Mathiness lets academic politics masquerade as science. Like mathematical theory, mathiness uses a mixture of words and symbols, but instead of making tight links, it leaves ample room for slippage between statements in the languages of words as opposed to symbols, and between statements with theoretical as opposed to empirical content. Because it is difficult to distinguish machines from mathematical theory, the market for lemons tells us that the market for mathematical theory might collapse, leaving only machines as entertainment that is worth little but cheap to produce.”
I use notebooks the whole time, normally in vs-code and with github copilot setup.
I found it quite painful to point it to a couple of environments I have, and confusing how i get it pointing to my gpt4 api keys. Once I did these two I was not sure how to prompt rather than typing a command.
Good luck with this, don't mean this in a critical way, just trying to give some feedback of what I think when I first try it.
Makes sense, i didn't think it was possible to maintain that complicated a codebase with so little changes.
Looking at the PR history, it's interesting how few PRs are being merged per year.
Of the ones that are they are very short, and typically buxfixes or changes to infrastructure rather than any new features.
I think I count 4 prs merged, with less than 20 lines of code altered, since 2022, and even going back until the beginning of the 2014 commit history, it's hard to find a PR that's altering core functionality.
https://github.com/overleaf/overleaf/pulls?q=is%3Apr+is%3Ame...
2024
+1 -1 A user should be created and a mail with an activation URL should be sent. #1208
+6 -6 Fix 502 errors due to IPv6 #1175
+1 -1 Make the number of max entities per project configurable #1108
2023
+6 -0 added SIGKILL timeouts for docker and phusion_image #1090
BigQuery has a generous 1TB/month free tier, and $6/tb afterwards. If you have small data, it's a pragmatic option, just make sure to use partitioning and sensible query patterns to limit the number of full-data scans, as you approach 'medium data' region.
There are some larger data-sizes, and query patterns, where either BigQuery Capacity compute pricing, or another vendor like Snowflake, becomes more economical.
https://cloud.google.com/bigquery/pricing
BigQuery offers a choice of two compute pricing models for running queries:
On-demand pricing (per TiB). With this pricing model, you are charged for the number of bytes processed by each query. The first 1 TiB of query data processed per month is free.
Queries (on-demand) - $6.25 per TiB - The first 1 TiB per month is free.Lendable | Senior Data-Scientist |Hybrid | UK, London | Full Time
Lendable have built the big three consumer finance products from scratch: loans, credit cards and car finance. Help us improve out best-in-class credit analytics that powers credit-decisioning.
https://jobs.eu.lever.co/lendable/0cb0ab21-0339-4ddd-84c3-39...