HNHacker News
TopNewBestAskShowJobs

abrichr

1,464 karma · joined November 16, 2011

Machine learning scientist, engineer, and entrepreneur.

- https://github.com/OpenAdaptAI/OpenAdapt - https://openadapt.ai - https://linkedin.com/in/richard-abrich - https://github.com/abrichr - https://MLDSAI.com - richard at openadapt dot ai

abrichr.at.hn

meet.hn/city/ca-Old-Toronto

Socials: - github.com/abrichr - x.com/abrichr - linkedin.com/in/richard-abrich

---

submissionscomments
abrichr··on FrontierMath was funded by OpenAI
I believe the paper being referenced is “Scaling Data-Constrained Language Models” (https://arxiv.org/abs/2305.16264).

For correctness, you can use a solver to verify generated data.

abrichr··on Skyvern Browser Agent 2.0: How We Reached State of the Art in Evals
This is exactly why in https://github.com/OpenAdaptAI/OpenAdapt:

1. Our goal is to automate repetitive tasks (not simple one off tasks that need human input, but repetitive tasks that human shouldn't be doing in the first place), and

2. The way we do this is by grounding in human demonstrations, so that the model has strict guardrails around what it can and can't do.

abrichr··on GPT-4o with scheduled tasks (jawbone) is available in beta
It seems there are some issues with the rollout.

Me:

> Give me positive feedback every hour

ChatGPT:

> Provide positive feedback

> Next run Jan 15, 2025

> Got it! I’ll send you positive feedback every hour.

An hour later, I received the following email:

```

Your scheduled task couldn't be completed

ChatGPT tried to complete Provide positive feedback multiple times, but it encountered an error and wasn't able to send. It will try again the next time this task is scheduled.

Open chat If you have any questions, please contact through the help center.

All the best, ChatGPT

```

abrichr··on 'Mainlined into UK's veins': Labour announces public rollout of AI
This is a bold plan, but there are real risks here. Opening NHS data for AI research sounds great in theory, but we’ve seen what happens when privacy safeguards fail. The Post Office scandal is a clear warning about relying too much on tech without accountability.

That said, the investment in sovereign compute capacity and miniature nuclear reactors is worth celebrating on its own. More nuclear energy is a win for everyone.

Let's hope they get it right.

abrichr··on AI founders will learn the bitter lesson
> No matter how good the AI gets, it can't answer about what it doesn't know. It can't perform a process for which it doesn't know the steps or the rules

This is exactly the motivation behind https://github.com/OpenAdaptAI/OpenAdapt: so that users can demonstrate their desktop workflows to AI models step by step (without worrying about their data being used by a corporation).

abrichr··on AI's next leap requires intimate access to your digital life
We designed https://github.com/OpenAdaptAI/OpenAdapt to ground the agent to human demonstrations, so that if you as a human don't engage in the dark pattern, neither will the agent.
abrichr··on AI's next leap requires intimate access to your digital life
The reason I created https://github.com/OpenAdaptAI/OpenAdapt is because I had the same concern, and wanted to create a way for users and organizations to avoid exfiltrating their data while still participating in the promise of AI.

I believe we can create a future in which you train the AI model that will assist you while participating in the profits it generates. "Build the future you want to live in."

abrichr··on New LLM optimization technique slashes memory costs
Looks like it is open source: https://github.com/FYYFU/HeadKV
abrichr··on Frontier Models are Capable of In-context Scheming
https://arxiv.org/abs/2412.04984

> Our findings demonstrate that frontier models now possess capabilities for basic in-context scheming [covertly pursuing misaligned goals], making the potential of AI agents to engage in scheming behavior a concrete rather than theoretical concern.

abrichr··on Show HN: Autotab – Programmable AI browser for turning web tasks into APIs
https://openadapt.ai is open source (MIT license).
abrichr··on Computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku
See https://github.com/OpenAdaptAI/OpenAdapt for an open source alternative that runs segmentation locally.
abrichr··on Computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku
See https://github.com/OpenAdaptAI/OpenAdapt for an open source alternative that includes a desktop app.
abrichr··on Computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku
> I expect people to be concerned about privacy implications of sending screenshots of the desktop

That's why in https://github.com/OpenAdaptAI/OpenAdapt we've built in several state-of-the-art PII/PHI scrubbers.

abrichr··on Computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku
We use operating system accessibility APIs when available in https://github.com/OpenAdaptAI/OpenAdapt.
abrichr··on Computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku
See https://github.com/OpenAdaptAI/OpenAdapt for an open source implementation that includes a desktop app OOTB.
abrichr··on Computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku
See https://github.com/OpenAdaptAI/OpenAdapt for an open source alternative that includes operating system accessibility API data and DOM information (along with bitmaps) where available.

We are also planning on extracting runtime information using COM/AppleScript: https://github.com/OpenAdaptAI/OpenAdapt/issues/873

abrichr··on Fraud, so much fraud
Most recent working archive.org snapshot:

https://web.archive.org/web/20240303093209/https://www.nia.n...

abrichr··on Apple Watch sleep apnea detection gets FDA approval
You can order them online, not sure if a prescription is required:

https://cpapoutlet.ca/

abrichr··on Ell: The language model programming library
This looks excellent. Thank you!
abrichr··on When A.I.'s Output Is a Threat to A.I. Itself
https://arxiv.org/abs/2408.15240

> We demonstrate that when using Gemma-based verifiers on algorithmic and grade-school math reasoning tasks, GenRM outperforms discriminative verifiers and LLM-as-a-Judge, showing a 16-64% improvement in the percentage of problems solved with Best-of-N. Furthermore, we show that GenRM scales favorably across dataset size, model capacity, and inference-time compute.

abrichr··on Artificial intelligence is losing hype
> I have an agent setup to focus on principles of Stoicism and its amazing how quickly I can get back on track just by having a short chat with it about how I'm feeling.

This sounds super useful. Can you please elaborate on the setup?

abrichr··on Show HN: PGlite – in-browser WASM Postgres with pgvector and live sync
As a workaround, maybe something like this could work:

db.js:

    const { PGlite } = require('@electric-sql/pglite');
    const db = new PGlite();

    async function execSQL(sql) {
        await db.exec(sql);
        return { success: true };
    }

    async function querySQL(sql, params) {
        const ret = await db.query(sql, params);
        return ret.rows;
    }

    // Command-line interface logic
    const action = process.argv[2];
    const sql = process.argv[3];
    const params = process.argv.slice(4);

    if (action === 'exec') {
        execSQL(sql).then(result => {
            console.log(JSON.stringify(result));
            process.exit();
        });
    } else if (action === 'query') {
        querySQL(sql, params).then(result => {
            console.log(JSON.stringify(result));
            process.exit();
        });
    }

main.py:

    import subprocess
    import json

    def execute_sql(sql: str) -> dict:
        result = subprocess.run(
            ['node', 'db.js', 'exec', sql],
            stdout=subprocess.PIPE,
            text=True
        )
        return json.loads(result.stdout)

    def query_sql(sql: str, params: list) -> list:
        result = subprocess.run(
            ['node', 'db.js', 'query', sql] + params,
            stdout=subprocess.PIPE,
            text=True
        )
        return json.loads(result.stdout)

    # Example Usage
    create_table_sql = """
    CREATE TABLE IF NOT EXISTS todo (
        id SERIAL PRIMARY KEY,
        task TEXT,
        done BOOLEAN DEFAULT false
    );
    """
    execute_sql(create_table_sql)

    query_result = query_sql("SELECT * FROM todo WHERE id = $1", ["1"])
    print(query_result)
abrichr··on Segment Anything Model and Friends
Example on Excel: https://x.com/OpenAdaptAI/status/1798502003045548480
abrichr··on Segment Anything Model and Friends
> Doesn’t work well for screen based content in general.

It's not perfect, but it works: https://github.com/OpenAdaptAI/OpenAdapt/pull/610

> the most recent latent space pod

Link: https://www.latent.space/p/sam2

abrichr··on Segment Anything Model and Friends
We are using Segment Anything Model at OpenAdapt for exactly this purpose: https://github.com/OpenAdaptAI/OpenAdapt/pull/610

It works surprisingly well despite the fact that the model was not trained on this type of data.

abrichr··on SAM 2: Segment Anything in Images and Videos
We use SAM to segment GUI elements in https://github.com/OpenAdaptAI/OpenAdapt
abrichr··on How large language models will disrupt data management [pdf]
https://gist.github.com/abrichr/455f0e569bf1bd104c696a7ad9e6...
abrichr··on XLSTMTime: Long-Term Time Series Forecasting with xLSTM
The paper links to their code:

https://github.com/muslehal/xLSTMTime

And their data, which includes daily exchange rates:

https://drive.google.com/drive/folders/1nuMUIADOc1BNN-uDO2N7...

abrichr··on Show HN: Find AI – Perplexity Meets LinkedIn
Thanks! Unfortunately same result
abrichr··on Show HN: Find AI – Perplexity Meets LinkedIn
> We're sorry, but something went wrong.

> If you are the application owner check the logs for more information.

Console:

cable_stream_source_element.js:22

       GET https://usefind.ai/searches/are-unhappy-with-robotic-process-automation-tools 500 (Internal Server Error)
← PreviousPage 2 of 20Next →