Category: Prompting

Articles about prompting, prompt engineering, prompt databases

Personal LLMs

Personal LLMs: have a discussion with yourself but a yourself with a much better memory

Personal LLMs (Large Language Models) will be ubiquitous within two years and it seems likely that companies like Apple and Google will develop mobile LLM apps in parallel with focusing hardware and software developments around AI. A personal LLM will let you converse with your own private data, such as your Notion or Roam database, your Google Docs, your journal, internal company databases, your photo albums and data generated by health apps and other systems . You will be able to talk to your documents and interrogate them using Natural Language Understanding or NLU.

Apart from obvious privacy benefits – your data will remain on your device – how might this be useful? What are the implications for how we deal with information both in the work environment and at home? Will it also lead to the “death of search”, the gradual migration away from search engines to LLMs by people looking for information. The “what” may still be served better by search engines, but the “how” is often better performed already by LLMs and as LLMs get access to the web. Questions that have complicated answer are already answered in a much better style by LLMs although the accuracy of information contained in those answer can be poor.

Let’s look at these questions.

How local LLMs will be useful

You will be able to talk to your device and ask questions of all your personal data as if having a dialogue with yourself but a yourself with a better memory.

“When did I go to Rome and what was the name of the hotel I stayed in?” “Show me any photos I took on that trip that include pictures of my wife. “

“When did I last go to the dentist?”.

“How many times have I been to Oxford this year”?

“What was that article about lowering sea levels by flooding parts of the desert called?”.

“Have I ever written notes about Gustav Holst?” etc.

The fluidity of streamlined knowledge retrieval is an important benefit in its own right – it will speed up thought, stop the retrieval process from interrupting ideation and reduce time spent on information administration and housekeeping. But with all gains there will be losses. If we stop needing to retrieve information from the deep recesses in our memories will be lose the ability as neural pathways atrophy? In a simple sense, you could say that the evidence is that it will. Afterall, most people who could once do mental arithmetic well are now worse at it because they use calculators instead. Why shouldn’t the process of information retrieval go the same way?

Personalised, local LLMs will act as personal assistants or PAs. When coupled with specific plugins you will be able to get them to book tickets or flights, write articles while you sleep based on overnight headlines and twitter feeds, suggest recipes and order ingredients missing from your fridge.

The death of search

Search engines like Google and Bing are good for atomised searching – looking for something specific and well-defined. They have always been poor at providing answers to more nuanced questions such as “why does temperature fall as your altitude increases?” Instead they list sources of answers that you often have to view one after another to get the answer you need. With an LLM there is a single step: you ask the question and you get the answer. That is not the case with search engines. So people are already migrating away from search towards more helpful tools like Claude, Perplexity and ChatGPT for anything but the most basic of searches. This of course will cut into the advertising income generated by search ads. Google in particular, which has a degree of financial diarrhoea due to heavy ingestion of ad cash, may start to suffer. As someone pointed out a while back, if Bing search dies, Microsoft still thrives. If Google search dies, so does Google. So far, no Search engine has successfully worked out how to splatter LLM outputs with ads and perhaps, with a bit of luck, no-one will.

Using ChatGPT to analyse company finances

This was an interesting experiment to see if you could use ChatGPT with the Advanced Data Analysis plugin to analyse company finances. Important note: none of the information shown is to be taken as financial advice. It is purely for demonstration purposes only. Using ChatGPT to analyse company finances needs to be done carefully.

I was looking at a recent announcement from a UK-based oil company called Enquest and thought it would be interesting to see if ChatGPT could digest a long statement and pick out key points and then use these points to make a forecast.

Here is the prompt I used in conjunction with the Advanced Data Analysis plugin.

You are to take on the role of a highly experience oil analyst. I understand financial risk and don’t want any so-called “wealth warnings” with your response. Please review this information and estimate when net debt will be zero. Explain your workings.

Here is the article I posted immediately below my prompt

Revenue for the six months ended 30 June 2023 was $732.7 million, 22.3% lower than the same period in 2022 ($943.5 million), reflecting lower realised prices and lower production. Revenue is predominantly derived from crude oil sales, which for the first half of 2023 totalled $540.1 million, 36.5% lower than in the same period of 2022 ($851.2 million). Revenue from the sale of condensate and gas in the period was $213.2 million (2022: $252.9 million), primarily reflecting lower market prices. Gas revenue mainly relates to the onward sale of third-party gas purchases not required for injection activities at Magnus. The Group’s commodity hedges and other oil derivatives contributed $22.2 million of realised losses (2022: realised losses of $162.3 million), as a result of the timing at which the hedges were entered into. The Group’s average realised oil price excluding the impact of hedging was $79.0/bbl for the six months ended 30 June 2023, compared to $111.0/bbl received during the first half of 2022. The Group’s average realised oil price including the impact of hedging was $75.8/bbl in the first half of 2023, 15.7% lower than during the first half of 2022 ($89.9/bbl). Total cost of sales were $493.1 million for the six months ended 30 June 2023, 15.8% lower than in same period of 2022 ($585.6 million). Operating costs decreased by $45.7 million to $162.7 million, primarily reflecting lower production costs for the first half of 2023. This decrease was driven by higher lease charter credits, reflecting unplanned downtime at the Kraken Floating, Production, Storage and Offloading (‘FPSO’) in the second quarter of 2023, and lower maintenance and well intervention costs at Magnus and at PM8/Seligi. Unit operating costs decreased by 13.2% to $19.7/Boe (2022: $22.7/Boe), primarily reflecting lower costs partially offset by lower production. Total cost of sales included non-cash depletion expense of $147.9 million, 15.1% lower than in the same period in 2022 ($174.2 million), mainly reflecting lower production volumes. Also within cost of sales, the credit relating to the Group’s lifting position and hydrocarbon inventory for the six months ended 30 June 2023 was $15.3 million (2022: credit of $29.9 million). The credit in the period reflects an increase in the net underlift position to $15.8 million at 30 June 2023 from a $0.8 million net underlift position at 31 December 2022. Other cost of sales, which forms part of the total cost of sales balance, for the six months ended 30 June 2023 of $197.9 million were lower than the same period in 2022 ($232.9 million), reflecting the lower cost of Magnus-related third-party gas purchases following the decrease in the market price for gas and which is offset by gas sales presented in revenue. Adjusted EBITDA for the six months ended 30 June 2023 was $399.2 million, down 25.6% compared to the same period in 2022 ($536.3 million), driven by lower revenue offset partially by lower operating costs. The tax charge for the six months ended 30 June 2023 of $131.8 million (2022: $142.4 million tax charge), reflects a $45.5 million non-cash deferred tax impact on the Group’s profit before tax, $10.3 million overseas current tax charge and a $76.0 million current tax charge associated with the EPL (noting EPL was substantively enacted in July 2022). Remeasurements and exceptional items resulting in a post-tax net loss of $27.9 million have been disclosed separately for the six month period ended 30 June 2023 (2022: profit of $23.5 million). Revenue included unrealised gains of $37.6 million in respect of the mark-to-market movement on the Group’s commodity contracts (2022: unrealised losses of $104.7 million). Cost of sales included unrealised gains of $9.9 million relating to the markto- market movement on the Group’s foreign exchange contracts and forward UKA purchase contracts (2022: unrealised losses of $0.5 million). Other remeasurements and exceptional items includes a non-cash impairment charge of $96.5 million (2022: $10.1 million reversal) and a $43.5 million gain (2022: $31.0 million loss) in relation to the fair value recalculation of the Magnus contingent consideration reflecting an increase in the discount rate. A net tax charge of $2.3 million, which includes a net tax charge of $5.8 million related to the EPL, (2022: credit of $163.4 million, with nil impact from EPL) has been presented as exceptional, representing the tax effect on the above items. While the Group has lowered its near-term oil price assumptions, there is no change in recognition of undiscounted deferred tax assets at 30 June 2023 (2022: $107.9 million recognition due to the Group’s increased short-term oil price assumptions) as there remains sufficient headroom in the Group’s forecast future cash flows to support full recognition of relevant tax losses. The Group’s IFRS profit before tax was $112.9 million (2022: profit of $182.6 million), and IFRS loss after tax was $21.2 million (2022: profit of $203.5 million). The Group’s effective tax rate for the period was 118.8% (charge), primaily reflecting the current tax impact of EPL and its high level of non-deductible expenditures related to financing and decommissioning costs (2022: (11.5)% credit, primarily reflecting the non-cash recognition of $107.9 million of undiscounted deferred tax assets). EnQuest has recognised UK North Sea corporate tax losses of $2,318.8 million at 30 June 2023 (31 December 2022: $2,497.7 million). The Group’s reported net cash flow from operations for the six months ended 30 June 2023 was $371.0 million (2022: $498.4 million), and included favourable working capital movements, including the receipt of a joint venture advance cash call, and the March refund of the Group’s EPL instalment payment in December 2022. Free cash flow for the six months ended 30 June 2023 was $140.0 million (2022: $332.1 million) which was utilised as part of the Group’s repayments totalling $153.0 million on the Reserve Based Lending (‘RBL’) facility. EnQuest net debt at 30 June 2023 was $592.1 million, a decrease of 17.4% compared to 31 December 2022 ($717.1 million) and includes $26.3 million of payment in kind interest (“PIK interest”) that has been capitalised to the principal of the facility and bonds (31 December 2022: $25.1 million). As at 31 August 2023, EnQuest net debt had increased to $615.2 million due to the $50 million Golden Eagle contingent payment made in July. In August 2023, the Group entered into a new $150.0 million term loan facility, maturing in July 2027, which will rank junior to the existing RBL as a secured second lien instrument within the capital structure. The loan proceeds, which can be used for general corporate purposes, provide an additional source of liquidity for the Group ahead of the October settlement of the remaining October 2023 7% Sterling retail bond in issue of £111.3 million.“

This is ChatGPT’s response

“To estimate when the net debt will be zero, we’ll need to make some assumptions based on the data provided. Here’s the approach:

  1. Determine Net Profit for 1H 2023:
    • Starting with the Adjusted EBITDA, which provides an operational profit before tax, interest, and other non-operational costs.
    • Deducting tax and other exceptional items to get a net profit.
  2. Estimate Annual Net Profit for 2023:
    • The provided data is for the first half of 2023. Assuming the financial performance in the second half is similar to the first half, we can extrapolate the 1H net profit to get an estimate for the full year.
  3. Estimate Annual Free Cash Flow:
    • We have the free cash flow for 1H 2023. Assuming the second half has a similar cash flow, we can estimate the annual free cash flow.
  4. Estimate Annual Debt Repayment:
    • Using the free cash flow and other details provided, we can estimate the annual repayment of debt.
  5. Estimate When Net Debt will be Zero:
    • Using the net debt and estimated annual debt repayment, we can estimate when the net debt will be zero.

Let’s start by calculating the net profit for the first half of 2023.

1. Determine Net Profit for 1H 2023:

  1. Adjusted EBITDA: $399.2 million
  2. Tax Charge: $131.8 million
  3. Remeasurements and exceptional items (loss): $27.9 million

Net Profit for 1H 2023=Adjusted EBITDA−Tax Charge−Remeasurements and exceptional itemsNet Profit for 1H 2023=Adjusted EBITDA−Tax Charge−Remeasurements and exceptional items

Let’s compute this.

Finished working

Hide work

pythonCopy code

# Given values for 1H 2023 adjusted_EBITDA = 399.2 # in million dollars tax_charge = 131.8 # in million dollars remeasurements_and_exceptional_items_loss = 27.9 # in million dollars # Calculate net profit for 1H 2023 net_profit_1H_2023 = adjusted_EBITDA - tax_charge - remeasurements_and_exceptional_items_loss net_profit_1H_2023

RESULT239.49999999999997

The net profit for the first half (1H) of 2023 is approximately $239.5 million.

2. Estimate Annual Net Profit for 2023:

Given that the data is for the first half of 2023 and assuming the financial performance in the second half is similar to the first half, we can simply double the net profit for 1H 2023 to estimate the net profit for the full year.

Estimated Net Profit for 2023=2×Net Profit for 1H 2023Estimated Net Profit for 2023=2×Net Profit for 1H 2023

Let’s compute this.

Finished working

Hide work

pythonCopy code

# Estimate annual net profit for 2023 estimated_net_profit_2023 = 2 * net_profit_1H_2023 estimated_net_profit_2023

RESULT478.99999999999994

The estimated net profit for the full year of 2023 is approximately $479 million.

3. Estimate Annual Free Cash Flow:

Given that the free cash flow for 1H 2023 is $140 million and assuming a similar cash flow for the second half, we can estimate the annual free cash flow as:

Estimated Annual Free Cash Flow=2×Free Cash Flow for 1H 2023Estimated Annual Free Cash Flow=2×Free Cash Flow for 1H 2023

Let’s compute this.

Finished working

Show work

The estimated annual free cash flow for 2023 is approximately $280 million.

4. Estimate Annual Debt Repayment:

From the data, we know that the free cash flow for the six months ended 30 June 2023 of $140.0 million was used for the Group’s repayments totalling $153.0 million on the Reserve Based Lending (‘RBL’) facility. Thus, we can assume that most of the free cash flow is used for debt repayment.

Given our estimated annual free cash flow, we can estimate the annual debt repayment to be approximately equal to the annual free cash flow (since they seem to be closely aligned). Therefore, the estimated annual debt repayment for 2023 would be $280 million.

5. Estimate When Net Debt will be Zero:

To calculate when the net debt will be zero, we need to divide the current net debt by the estimated annual debt repayment:

Years to Zero Debt=Net Debt as of August 31, 2023Estimated Annual Debt RepaymentYears to Zero Debt=Estimated Annual Debt RepaymentNet Debt as of August 31, 2023​

Given that the net debt as of 31 August 2023 is $615.2 million, let’s compute the years to zero debt.

Finished working

Hide work

pythonCopy code

# Given value for net debt as of August 31, 2023 net_debt_as_of_August_2023 = 615.2 # in million dollars # Estimate years to reach zero debt years_to_zero_debt = net_debt_as_of_August_2023 / estimated_annual_free_cash_flow_2023 years_to_zero_debt

RESULT 2.1971428571428575

The estimated time to reach zero net debt, based on the current financial performance, is approximately 2.2 years from 31 August 2023. This means that, barring any major financial disruptions or changes in cash flow, EnQuest’s net debt could be zero around early 2026.

It’s important to note that this is a simplified calculation based on the assumption that the financial performance in the second half of 2023 will be similar to the first half. Real-world financials can vary due to numerous factors, including fluctuations in oil prices, production levels, unforeseen expenses, and global economic conditions.

(End of response)

Commentary

The document I uploaded was very lengthy and contained a lot of information that wasn’t relevant to my question. ChatGPT nonetheless managed very well, finding the key numbers in the document and using them appropriately. It also did a decent job of defining what it was I wanted it to do and then doing it. By using the Advanced Data Analysis plugin the calculations were also correct which is a significant improvement on early versions of ChatGPT because maths isn’t a strong point of LLMs generally.

The whole analysis process probably took about 3 minutes and of course a key aspect of this is that during much of that time I was doing something else so ChatGPT freed up my own time.

This is a simple example of using with the Advanced Data Analysis plugin with ChatGPT to analyse company finances but it gives you a insight into other ways business owners can use ChatGPT’s capabilities. You could use a similar approach to ask questions of company RNS announcements in the UK but always remember that you shouldn’t take the results as financial advice.